METHOD FOR INTEGRATING IMAGE ANALYSIS, LONGITUDINAL TRACKING OF A REGION OF INTEREST AND UPDATING OF A KNOWLEDGE REPRESENTATION

FIELD OF THE DISCLOSURE

The present disclosure generally relates to image interpretation, and more particularly, to systems and methods for generating image reports.

BACKGROUND OF THE DISCLOSURE

In current image interpretation practice, such as diagnostic radiology, a specialist trained in interpreting images and recognizing abnormalities may look at an image or an image sequence on an image display and report any visual findings by dictating or typing the findings into a report template. The dictating or typing usually includes a narration of the finding, a description about the location of the visual phenomena, abnormality, or region of interest within the images being reported on. The recipient of the report is often left to further analyze the contents of the narrative text report without having easy access to the underlying image. More particularly, in current reporting practice, there is no link between the specific location in the image and the finding associated with the visual phenomena, abnormality, or region of interest, in the image. A specialist also may have to compare a current image with an image and report previously done. This leaves the interpreter to refer back and forth between the image and the report.

Computer-aided detection (CAD) systems are known in the art and are usually confined to detecting and classifying conspicuous structures in the image data. Computer-aided diagnosis (CAD) systems are used in mammography to highlight micro calcification clusters and hyperdense structures in the soft tissue. Computer-aided simple triage (CAST) is another type of CAD, which performs a fully automatic initial interpretation and triage of studies into some meaningful categories (e.g. negative and positive). Unfortunately, these prior art systems are limited to describing the location of the visual phenomena within the image file. By manner of illustration, the coordinate system provided by the CAD system cannot be used to guide a biopsy needle because it fails to identify the relative position within the organ or sample structure.

While such inconveniences may pose a seemingly insignificant risk of error, a typical specialist must interpret a substantial amount of such images in short periods of time, which further compounds the specialist's fatigue and vulnerability to oversights. This is especially critical when the images to be interpreted are medical images of patients with their health being at risk.

General articulation and narration of an image interpretation may be facilitated with reference to structured reporting templates or knowledge representations. One example of a knowledge representation in the form of a semantic network is the Systematized Nomenclature of Medicine-Clinical Terms (SNOMED-CT), which is a systematically organized and computer processable collection of medical terminology covering most areas of clinical information, such as diseases, findings, procedures, microorganisms, pharmaceuticals, and the like. SNOMED-CT provides a consistent way to index, store, retrieve, and aggregate clinical data across various specialties and sites of care. SNOMED-CT also helps in organizing the content of medical records, and in reducing the inconsistencies in the way data is captured, communicated, encoded, and used for clinical care of patients and research.

Another example is the Breast Imaging-Reporting and Data System (BI-RADS), which is a quality assurance tool originally designed for use with mammography. Yet another example is RadLex, a lexicon for uniform indexing and retrieval of radiology information resources, which currently includes more than 30,000 terms. Applications include radiology decision support, reporting tools and search applications for radiology research and education. Reporting templates developed by the Radiological Society of North America (RSNA) Reporting Committee use RadLex terms in their content. Reports using RadLex terms are clearer and more consistent, reducing the potential for error and confusion. RadLex includes other lexicons and semantic networks, such as SNOMED-CT, BI-RADS, as well as any other system or combination of systems developed to help structure and standardize reporting. Richer forms of semantic networks in terms of knowledge representation are ontologies. Knowledge representations may also include probability models and identifying characteristics from image data generated by image segmentation and classification algorithms. Ontologies are encoded using ontology languages and commonly include the following components: instances (the basic or “ground level” objects), classes (sets, collections, concepts, classes in programming, types of objects, or kinds of things), attributes (aspects, properties, features, characteristics, or parameters that objects), relations (ways in which classes and individuals can be related to one another), function terms (complex structures formed from certain relations that can be used in place of an individual term in a statement), restrictions (formally stated descriptions of what must be true in order for some assertion to be accepted as input), rules (statements in the form of an if-then sentence that describe the logical inferences that can be drawn from an assertion in a particular form, axioms (assertions, including rules, in a logical form that together comprise the overall theory that the ontology describes in its domain of application), and events (the changing of attributes or relations).

Currently existing image reporting mechanisms do not take full advantage of knowledge representations to assist interpretation while automating reporting. In particular, currently existing systems are not fully integrated with knowledge representations to provide seamless and effortless reference to knowledge representations during articulation of findings. Additionally, in order for such a knowledge representation interface to be effective, there must be a brokering service between the various forms of standards and knowledge representations that constantly evolve. While there is a general lack of such brokering service between the entities of most domains, there is an even greater deficiency in the available means to promote common agreements between terminologies, especially in image reporting applications. Furthermore, due to the lack of more streamlined agreements (alignment) between knowledge representations in image reporting, currently existing systems also lack means for automatically tracking the development of specific and related cases for inconsistencies or errors so that the knowledge representations may be updated to provide more accurate information in subsequent cases. Such tracking means provide the basis for a probability model for knowledge representations.

In light of the foregoing, there is a need for an improved system and method for integrating image analysis, longitudinal tracking of a region of interest and updating of a knowledge representation.

SUMMARY OF THE DISCLOSURE

Disclosed is a method for integrating image analysis, longitudinal tracking of a region of interest and updating of a knowledge representation, said method comprising the steps of:

retrieving an image representation of a sample structure from an image database;

automatically selecting a generic structure from a database containing a plurality of generic structures based on an imaging modality of the sample structure, at least one knowledge representation stored in a second database, said knowledge representation associated with said selected generic structure, the knowledge representation being specific to the imaging modality;

mapping the selected generic structure to the sample structure;

automatically determining at least one region of interest within the sample structure or allowing the user to select a region of interest;

automatically selecting at least one diagnostic finding or allowing the user to select at least one diagnostic finding from a focused set knowledge representations;

retrievably storing the at least one diagnostic finding in an electronic record; and

monitoring the electronic record for changes to the at least one diagnostic finding or new diagnostic findings and using such changes or new diagnostic findings to update the knowledge representation in the second database.

The aforementioned method wherein the step of allowing the user to select at least one diagnostic finding from the focused set of knowledge representation includes allowing the user to enter the at least one diagnostic finding using free-form text.

The aforementioned method wherein the selected generic structure is related to the sample structure by imaging modality and one or more attributes selected from the group (size, dimensions, area, shape, volume, weight, density, location, anatomical organ, and orientation).

The aforementioned method further, wherein the selected generic structure has coordinate data defined therein.

The aforementioned method, further comprising:

using the coordinate data to generate natural language statements describing a location of the region of interest in the anatomy;

automatically generating a diagnostic report based on the at least one diagnostic finding, and including the natural language statements describing the location in the anatomy of the region of interest; and

storing the diagnostic report in the electronic record.

The aforementioned method, wherein the knowledge representation is specific to an anatomical organ in which the region of interest is located and the imaging modality.

The aforementioned method further comprising:

automatically generating a diagnostic report based on the selections or free-form text entries, and including natural language statements describing the location in the anatomy of the region of interest; and storing the diagnostic report in the electronic record.

The aforementioned method, wherein the step of automatically selecting a generic structure from among a plurality of generic structures is based on the imaging modality and a comparison of content of the sample structure to the content of the generic structure.

The aforementioned method further comprising:

for each at least one region of interest automatically selecting follow-up care or prompting the user to select from a focused set of follow-up care options; and storing the selected follow-up care in the electronic record.

The aforementioned method, wherein the step of monitoring the electronic record includes checking for changes to the selected follow-up care and using such changes to update the knowledge representation in the second database.

The aforementioned method, wherein the step of monitoring the electronic record includes checking for changes to treatment outcome and using such changes to update the knowledge representation in the second database.

Also disclosed is a method for integrating image analysis, longitudinal tracking of a region of interest and updating of a knowledge representation, comprising the steps of:

retrieving an image representation of sample structure depicting at least a portion of an anatomical organ from an image database;

determining at least one region of interest within the sample structure or allowing the user to select a region of interest;

automatically selecting at least one diagnostic finding or allowing the user to select at least one diagnostic finding from a focused set of knowledge representations specific to the anatomical organ and an imaging modality used to capture the image representation;

retrievably storing the at least one diagnostic finding in the electronic record;

monitoring the electronic record for changes and/or additions to the at least one diagnostic finding and updating the knowledge representation to reflect the changes and/or additions to the at least one diagnostic finding.

Also disclosed is a method for integrating image analysis, longitudinal tracking of a region of interest and updating of a knowledge representation, the method comprising the steps of:

recording at least one diagnostic finding for a given region of interest in an image database;

monitoring the electronic record for changes to the at least one diagnostic finding for the region of interest; and

automatically updating a knowledge representation stored in a database to reflect the changes to the at least one diagnostic finding for the region of interest.

The aforementioned method, further comprising:

retrieving an image representation of sample structure depicting at least a portion of an anatomical organ from an image database;

automatically determining at least one region of interest within the sample structure or allowing the user to select a region of interest;

automatically selecting at least one diagnostic finding or allowing the user to select at least one diagnostic finding from a focused set of knowledge representations specific to at least one of the anatomical organ and an imaging modality used to capture the image representation; and

retrievably storing the at least one diagnostic finding in the electronic record.

These and other aspects of this disclosure will become more readily apparent upon reading the following detailed description when taken in conjunction with the accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a diagrammatic view of an exemplary system for supporting an image reporting method;

FIG. 2 is a schematic view of an exemplary image reporting device constructed in accordance with the teachings of the disclosure;

FIG. 3 is a diagrammatic view of an exemplary algorithm for image reporting;

FIGS. 4A-4B are diagrammatic views of a sample structure;

FIG. 5 is a diagrammatic view of a three-dimensional mapping technique;

FIG. 6 is a diagrammatic view of a related three-dimensional mapping technique;

FIGS. 7A-7C are diagrammatic views of a generic structure;

FIGS. 8A-8D are illustrative views of a warping process;

FIGS. 9A-9B are diagrammatic views of another sample structure;

FIGS. 10A-10B are diagrammatic views of yet another sample structure;

FIGS. 11A-11C are diagrammatic views of exemplary image reports;

FIG. 12 is a schematic view of an exemplary image reporting system integrated with knowledge representation systems;

FIGS. 13A and 13B is another diagrammatic view of a sample structure also illustrating knowledge representations;

FIG. 14 is a diagrammatic view of showing the user being prompted to select from one of the guidelines or enter a user specified instruction for follow-up care;

FIG. 15 is a diagrammatic view of a knowledge base or ontology showing recommended treatments for each of a plurality of diagnoses;

FIG. 16 is a diagrammatic view of a sample structure showing a region of interest;

FIG. 17 is a table providing both current and historical information regarding a region of interest;

FIGS. 18-20 are flow diagrams of a method for longitudinally tracking changes to diagnostic findings; and

FIGS. 21-22 are diagrams depicting the utilization of neural networks.

While the present disclosure is susceptible to various modifications and alternative constructions, certain illustrative embodiments thereof have been shown in the drawings and will be described below in detail. It should be understood, however, that there is no intention to limit the present invention to the specific forms disclosed, but on the contrary, the intention is to cover all modifications, alternative constructions, and equivalents falling with the spirit and scope of the present invention.

DETAILED DESCRIPTION

Referring now to FIG. 1, an exemplary system 100 within which an image interpretation and reporting method may be integrated is provided. As shown, the system 100 may include a central network 102 by which different components of the system 100 may communicate. For example, the network 102 may take the form of a wired and/or wireless local area network (LAN), a wide area network (WAN), such as the Internet, a wireless local area network (WLAN), a storage or server area network (SAN), and the like. The system 100 may also include image capture devices 104 configured to capture or generate two-dimensional and/or three-dimensional images. In medical imaging, for example, the image capture devices 104 may include one or more of a mammography device, a computed tomography (CT) device, an ultrasound device, an X-ray device, a fluoroscopy device, a film printer, a film digitizer, and the like. One or more images of a sample structure captured by the image capture devices 104 may be transmitted to an image server 106 and/or an image database 108 directly or through a network 102.

The image server 106, image database 108 and/or network 102 of FIG. 1 may be configured to manage the overall storage, retrieval and transfer of images, as in Picture Archiving and Communication System (PACS) in accordance with Digital Imaging and Communications in Medicine (DICOM) standards, for example. In medical applications, each medical image stored in the DICOM database may include, for instance, a header containing relevant information, such as the patient name, the patient identification number, the image type, the scan type, or any other classification type by which the image may be retrieved. Based on the classification type, the server 106 may determine where and how specific images are stored, associate the images with any additional information required for recalling the images, sort the images according to relevant categories and manage user access to those images. In further alternatives, the storage, retrieval and transfer of images may be managed and maintained within the network 102 itself so as to enable services, for example, in an open source platform for individual users from any node with access the network 102. In an application related to medical imaging, for example, each medical image may be tied to a particular patient, physician, symptom, diagnosis, or the like. The stored images may then be selectively recalled or retrieved at a host 110.

As shown in FIG. 1, one or more hosts 110 may be provided within the system 100 and configured to communicate with other nodes of the system 100 via the network 102. Specifically, users with appropriate authorization may connect to the image server 106 and/or image database 108 via the network 102 to access the images stored within the image database 108. In medical applications, for example, a host 110 may be used by a physician, a patient, a radiologist, or any other user granted access thereto. In alternative embodiments, the system 100 may be incorporated into a more localized configuration wherein the host 110 may be in direct communication with one or more image capture devices 104 and/or an image database 108.

Turning now to FIG. 2, one exemplary image reporting device 200 as applied at a host 110 is provided. The image reporting device 200 may essentially include a computational device 202 and a user interface 204 providing user access to the computational device 202. The user interface 204 may include at least one input device 206 which provides, for example, one or more of a keypad, a keyboard, a pointing device, a microphone, a camera, a touch screen, or any other suitable device for receiving user input. The user interface 204 may further include at least one output or viewing device 208, such as a monitor, screen, projector, touch screen, printer, or any other suitable device for outputting information to a user. Each of the input device 206 and the viewing device 208 may be configured to communicate with the computational device 202.

In the particular image reporting device 200 of FIG. 2, the computational device 202 may include at least one controller or microprocessor 210 and a storage device or memory 212 configured to perform image interpretation and/or reporting. More specifically, the memory 212 may be configured to at least one algorithm for performing the image reporting function, while the microprocessor 210 may be configured to execute computations and actions for performing according to the stored algorithm. In alternative embodiments, the microprocessor 210 may include on-board memory 213 similarly capable of storing the algorithm and allowing the microprocessor 210 access thereto. The algorithm may also be provided on a removable computer-readable medium 214 in the form of a computer program product. Specifically, the algorithm may be stored on the removable medium 214 as control logic or a set of program codes which configure the computational device 202 to perform according to the algorithm. The removable medium 214 may be provided as, for example, a compact disc (CD), a floppy, a removable hard drive, a universal serial bus (USB) drive, a flash drive, or any other form of computer-readable removable storage.

Still referring to FIG. 2, the image reporting device 200 may be configured such that the computational device 202 is in communication with at least one image source 216. The image source 216 may include, for example, an image capture device 104 and/or a database of retrievable images, as shown in FIG. 1. In a localized configuration, the computational device 202 may be in direct wired or wireless communication with the image source 216. In still other alternatives, the image source 216 may be established within the memory 212 of the computational device 202. In a network configuration, the computational device 202 may be provided with an optional network or communications device 218 so as to enable a connection to the image source 216 via a network 102.

As shown in FIG. 3, a flow diagram of an exemplary algorithm 300 by which an image reporting device 200 may conduct an image reporting session is provided. In an initial step 302, one or more images of a sample structure to be interpreted may be captured and/or recorded. The images may include, for instance, one or more two-dimensional medical images, one or more three-dimensional medical images, or any combination thereof. The sample structure to be interpreted may be, for instance, a patient, a part of the anatomy of a patient, or the like. More specifically, in an image reporting session for medical applications, the images that are captured and/or recorded in step 302 may pertain to a mammography screening, a computer tomography (CT) scan, an ultrasound, an X-ray, a fluoroscopy, or the like.

In an optional step 304, the captured or recorded images may be copied and retrievably stored at an image server 106, an image database 108, a local host 110, or any other suitable image source 216. Each of the copied and stored images may be associated with information linking the images to a sample subject or structure to be interpreted. For instance, medical images of a particular patient may be associated with the patient's identity, medical history, diagnostic information, or any other such relevant information. Such classification of images may allow a user to more easily select and retrieve certain images according to a desired area of interest, as in related step 306. For example, a physician requiring a mammographic image of a patient for the purposes of diagnosing breast cancer may retrieve the images by querying the patient's information via one of the input devices 206. In a related example, a physician conducting a case study of particular areas of the breast may retrieve a plurality of mammographic images belonging to a plurality of patients by querying the image server 106 and/or database 108 for those particular areas.

Upon selecting a particular study in step 306, one or more retrieved images may be displayed at the viewing device 208 of the image reporting device 200 for viewing by the user as in step 308. In alternative embodiments, for example, wherein the image source 216 or capture device 104 is local to the host 110, steps 304 and 306 may be omitted and recorded images may be displayed directly without copying the images to an image database 108.

Exemplary images 310 that may be presented at the viewing device 208 are provided in FIGS. 4A-4B. The views contained in each of FIGS. 4A-4B may be simultaneously presented at a single display of a viewing device 208 to the reader so as to facilitate the reader's examination and comprehension of the underlying anatomical object. Alternatively, one or more components or views within each of FIGS. 4A-4B may also be provided as individual views that are simultaneously and/or sequentially presentable at multiple displays of the viewing device 208. The images 310 may include one or more two-dimensional (2D) or three-dimensional (3D) views of an image representation of an image 312 to be interpreted. In the particular views of FIGS. 4A-4B, two-dimensional medical image representations or mammographic images 310 of a breast 312 are provided. Moreover, the displays of FIGS. 4A-4B may include the right mediolateral oblique (RMLO) view of the sample breast 312, as well as the right craniocaudal (RCC) view of the corresponding sample breast 312. Alternatively, one or more three-dimensional views of a sample breast structure 312 may be displayed at the viewing device 208 of the image interpretation and reporting device 200.

Additionally, the images 310 may also provide views of an image representation of a reference structure 314 for comparison. The reference structure 314 may be any one of a prior view of the sample structure 312, a view of a generic structure related to the sample structure 312, a benchmark view of the sample structure 312, or the like.

The generic structure may be related to the sample structure by imaging modality. The generic structure may further be related to the generic structure by one or more attributes including size, dimensions, area, shape, volume, weight, density, location, anatomical organ, and orientation.

The selected generic structure may have coordinate data defined therein. As will be explain in further detail below, the coordinate data may be used in describing the location of the region of interest in the anatomy. The system automatically selects a generic structure from among a plurality of generic structures based on the imaging modality. The system may further select the generic structure based on a comparison of content of the sample structure to the content of the generic structure.

The images 310 may even be provided using different imaging modalities such as computer tomography (CT) scan, an ultrasound, an X-ray, a fluoroscopy, or the like. These different imaging modalities may be linked using image registration techniques commonly known in the art. For the sake of clarity, the term registration as used herein refers to known techniques for correlating a point or a region of interest in a first image with the corresponding location or region in a second image. It should be appreciated that the term registration applies whether images are both from the same imaging modality or if the images were captured using different imaging modalities. Furthermore, the reference structure 314 may be automatically selected and supplied by the image reporting device 200 in response to the sample structure 312 that is retrieved. The image reporting device 200 may prompt the user to confirm that the appropriate reference structure 314 was selected. Moreover, based on certain features of the sample structure 312 in question, the image reporting device 200 may automatically retrieve a comparable reference structure 314 from a collection of reference structures 314 stored at an image source 216, image database 108, or the like. Alternatively, a user may manually select and retrieve a comparable reference structure 314 for viewing.

Although some retrieved image representations of sample structures 312 may already be in three-dimensional form, many retrieved image representations of a sample structure 312 may only be retrievable in two-dimensional form. Accordingly, the step 308 of displaying an image representation of a sample structure 312 may further perform a mapping sequence so as to reconstruct and display a three-dimensional image representation of the sample structure 312 using any one of a variety of known mapping techniques. As shown in FIG. 5, for example, a computer tomography (CT) image representation of a sample structure 316 of a human head may be retrieved as a collection of two-dimensional images 318, wherein each image 318 may display one lateral cross-sectional view of the sample head structure 316. In such a case, the individual cross-sectional images 318 may be combined to reconstruct the three-dimensional head structure 316 shown. Such mapping techniques may be extended to reconstruct a three-dimensional representation of a complete human anatomy as one sample structure 312. Other known techniques for mapping, as demonstrated in FIG. 6 for example, may exist, wherein a deformable mesh 320 laid over a known data distribution may define the geometric transformation to a three-dimensional structure 322 of unknown data distribution after several iterations of local registrations. Additional mapping techniques may be used in which the deformation of a three-dimensional structure may be represented by a three-dimensional grid, for example, composed of tetraeders, or with three-dimensional radial basis functions. Depending on the resolution applied, the interior content of a three-dimensional image may be well-defined and segmented so as to be automatically discernable by software, for instance. For medical image interpretation practices, such voxel data and the resulting three-dimensional contents may be used to represent and distinguish between any underlying tissues, organs, bones, or the like, of a three-dimensional part of the human anatomy. Still further refinements for mapping may be applied according to, for instance, Hans Lamecker, Thomas Hermann Wenckebach, Hans-Christian Hege. Atlas-based 3D-shape reconstruction from x-ray images. Proc. Int. Conf. of Pattern Recognition (ICPR2006), volume I, p. 371-374, 2006, wherein commonly observed two-dimensional images may be processed and morphed according to a known three-dimensional model thereof so as to reconstruct a refined three-dimensional representation of the image initially observed.

In a similar manner, the algorithm 300 may map a generic structure 324, as shown in FIGS. 7A-7C, to the sample structure 312 of FIGS. 4A-4B. A generic structure 324 may include any known or well-defined structure that is related to the sample structure 312 and/or comparable to the sample structure 312 in terms of size, dimensions, area, volume, weight, density, orientation, or other relevant attributes. Thus, the generic structure may be a prior image of the same structure thereby enabling longitudinal comparison of the region of interest. The generic structure 324 may also be associated with known coordinate data. Coordinate data may include pixel data, bitmap data, three-dimensional data, voxel data, or any other data type or combinations of data suitable for mapping a known structure onto a sample structure 312. For example, the embodiments of FIGS. 7A-7C illustrate an image representation of a generic breast structure 324 that is comparable in size and orientation to the corresponding sample breast structure 312, and further, includes coordinate data associated therewith. Moreover, in the mammographic images 310 of FIGS. 7A-7C, the coordinate data may be defined according to a coordinate system that is commonly shared by any sample breast structure 312 and sufficient for reconstructing a three-dimensional image, model or structure thereof. By mapping or overlaying the coordinate data of the generic structure 324 onto the sample structure 312, the image reporting algorithm 300 may be enabled to spatially define commonly shared regions within the sample structure 312, and thus, facilitate any further interpretations and/or annotations thereof. By mapping, for instance, the generic breast structure 324 of FIGS. 7A-7C to the sample structure 312 of FIGS. 4A-4B, the algorithm 300 may be able to distinguish, for example, the superior, inferior, posterior, middle, and anterior sections of the sample breast structure 312 as well as the respective clock positions.

As will be described below in further detail, different taxonomies are associated with each generic structure. Thus, the selection of a given generic structure restricts the universe of applicable taxonomies. Moreover, different taxonomies are associated with each imaging modality. The taxonomy used to describe a computer tomography (CT) scan of a sample structure is different from the taxonomy used to describe an ultrasound image of the same sample structure. Likewise, X-ray, a fluoroscopy, or the like each use their own unique taxonomy. The image reporting system of the present invention selects the appropriate taxonomy based on the imaging modality and the generic structure thereby facilitating ease of use and ensuring consistent usage of terminology in the reports.

As with reference structures 314, selection of a compatible generic structure 324 may be automated by the image reporting device 200 and/or the algorithm 300 implemented therein. Specifically, an image database 108 may comprise a knowledgebase of previously mapped and stored sample structures 312 of various categories from which a best-fit structure may be designated as the generic structure 324 for a particular study. The knowledge representation may be stored within the knowledgebase.

As used in the present disclosure, the term “knowledge representation” includes identifying characteristics of biological structures and knowledge about visual representation of normal and abnormal tissue. The term “tissue” includes both bone and soft tissue, i.e., any biological structure. The term “knowledge representation” also includes genetic data, demographic data, effectiveness of treatments, behavioral data, nutritional data, i.e., any health-related data.

Knowledge representations include identifying characteristics from annotated region of interests and the tracking of changes to the medical records related to the region of interest and the treatment outcomes. A preferred embodiment of the knowledge representation includes computer vision and machine learning frameworks such as the open-source software library TensorFlow, more specifically artificial convolutional neural networks to advance the knowledge representation with knowledge of identifying characteristics within image data. A convolutional neural network is trained with an initial data set as depicted in FIG. 21. The input data includes annotated image files typically in DICOM format, pathology results, health records, genetic data, behavioral data etc. The image data consists of positive space (pixel data within the region of interest) and associated finding and negative space (pixel data outside of the region of interest) and associated general findings. Non-image date such as pathology results, health records, genetic data, behavioral data etc. contain information on the accuracy of the diagnostic finding and effectiveness of the treatment recommendation. The training data may be fed into the learning process one at a time or as a batch. As such training is computationally demanding, a distributed approach depicted in FIG. 22 may be utilized.

Distributed Training of the Network as Part of the Knowledge Representation (FIG. 22)

The knowledge of identifying characteristics within image data is used to automatically select regions of interest and automatically select a diagnostic finding for such region as part of the diagnostic process.

The accuracy of the knowledge representation is continuously improved by means of online machine learning methods in which data becomes available in a sequential order and is used to update our best predictor for future data at each step, as opposed to batch learning techniques which generate the best predictor by learning on the entire training data set at once. This method of progressive incremental learning is independent of the number of class constraints and it can learn new classes while still retaining the knowledge of previous classes. Whenever a new class (non-native to the knowledge learned thus far) is encountered, the classifier gets remodeled automatically and the parameters are calculated in such a way that it retains the knowledge learned thus far.

As the quality of images continuously improves and new imaging modalities emerge, the preferred embodiment ages older data by automatically assigning a lower weight to older training images whereas newer data is automatically assigned a higher weight.

The preferred embodiment includes a sequence of machine learning models ML_t, where ML_{t+1} is trained later in time than ML_t, each model ML_t trained based on a set of training data D_t consisting of training samples s_{t,i} with respective training weights w_{t,i}. Each sample s_{t,i} that is similar to a sample s having a reduced weight w_{t,i}<w_{t−1, k}, and samples with updated outcome having an increased weight w_{t,i}>w_{t−1, k}.

Diagnostic findings which are verified by non-image data such as pathology results is automatically assigned the highest weight. In one alternative, an approximated generic structure 324 may be constructed based on an average of attributes of all previously mapped and stored sample structures 312 relating to the study in question. Accordingly, the ability of the algorithm 300 to approximate a given sample structure 312 may improve with every successive iteration. Alternatively, a user may manually filter through an image source 216 and/or an image database 108 to retrieve a comparable generic structure 324 for viewing.

Referring back to the algorithm 300 of FIG. 3, once the image representations of the sample structure 312 are mapped and displayed in step 308, the algorithm 300 may enable selection of one or more points or regions of interest (ROIs) within the image representation of the sample structure 312 in step 328. As illustrated in FIGS. 4A-4B, a visual phenomena, abnormality, or region of interest 326 may be determined based on the contents of the image representation of the sample structure 312. For example, in the mammographic images 310 of FIGS. 4A-4B, a region of interest 326 may correspond to a plurality of calcifications disposed within the sample breast structure 312. Such a region of interest 326 may be determined manually by a user viewing the sample structure 312 from an image reporting device 200. One or more regions of interest 326 may also be automatically located by the image reporting algorithm 300. For example, the algorithm 300 may automatically and/or mathematically compare contents of the image representation of the sample structure 312 with the contents of image representation of the reference structure 314, as shown in FIGS. 4A-4B. In some embodiments, the algorithm 300 may similarly enable recognition of contents within an image representation of a generic structure 324.

During such comparisons, it may be beneficial to provide comparison views between a sample structure 312 and a reference structure 314, as demonstrated in FIGS. 4A-4B. However, not all image representations of the reference structure 314 may be retrieved in an orientation that is comparable to that of the sample structure 312, as shown in FIGS. 8A-8D. Accordingly, the algorithm 300 may be configured to automatically warp the position, orientation and/or scale of the image representation of the reference structure 314 to substantially match that of the sample structure 312. In alternative embodiments, the algorithm 300 may be configured to automatically warp the image representation of the sample structure 312 to that of the reference structure 314.

In an exemplary warping process, the algorithm 300 may initially determine two or more landmarks 330, 332 that are commonly shared by the sample and reference structures 312, 314. For example, in the mammographic images 310 of FIGS. 8A-8D, the first landmark 330 may be defined as the nipple of the respective breast structures 312, 314, while the second landmark 332 may be defined as the pectoralis major muscle line. Forming an orthogonal baseline 334 from the first landmark 330 to the second landmark 332 of each structure 312, 314 may provide a basis from which the algorithm 300 may determine the spatial offset that needs to be adjusted. Based on the coordinate mapping performed earlier in step 308 and the detected differences between the respective landmark positions, the algorithm 300 may automatically adjust, rotate, shift, scale or warp one or both of the sample structure 312 and the reference structure 314 to minimize the offset. For instance, in the example of FIGS. 8A-8D, the algorithm 300 may rotate the image representation of the prior reference structure 314 in the direction indicated by arrow 336 until the orientations of the respective landmark baselines 334 are substantially parallel. In an alternative embodiment, the generic structure 314 may be substituted for the reference structure 314, in which case similar warping processes may be employed to minimize any skewing of views.

Still referring to step 328 of FIG. 3, once at least one region of interest 326 has been determined, the algorithm 300 may further link the region of interest 326 with the coordinate data that was mapped to the sample structure 312 during step 308. Such mapping may enable the algorithm 300 to define the spatial location of the region of interest 326 with respect to the generic structure 314 and not only with respect to the view or image representation of the sample structure 312 shown. Moreover, the algorithm 300 may be able to at least partially track the location of the region of interest 326 within the sample structure 312 regardless of the view, position, orientation or scale of the sample structure 312. In particular, if the algorithm 300 is configured to provide multiple views of a sample structure 312, as in the mammographic views of FIGS. 4A-4B for example, step 340 of the algorithm 300 may further provide a range or band of interest 338 in one or more related views corresponding to the region of interest 326 initially established. Based on manual input from a user or automated recognition techniques, step 342 of the algorithm 300 may then determine the corresponding region of interest 326 from within the band of interest 338.

As in the warping techniques previously discussed, in order to perform the tracking steps 340 and 342 of FIG. 3, the algorithm 300 may identify at least two landmarks 330, 332 within the sample structure 312 in question. In the mammographic views of FIGS. 9A-9B shown, for example, the first landmark 330 may be defined as the nipple, and the second landmark 332 may be defined as the pectoralis major muscle line. The algorithm 300 may then define a baseline 334a as, for example, an orthogonal line extending from the nipple 330 to the pectoralis major muscle line 332. As demonstrated in FIG. 9A, a user may select the region of interest 326a on the right mediolateral oblique (RMLO) view of the sample breast structure 312. After the selection, the algorithm 300 may project the region of interest 326a onto the baseline 334a, from which the algorithm 300 may then determine a first distance 344a and a second distance 346. The first distance 344a may be determined by the depth from the first landmark 330 to the point of projection of the region of interest 326a on the baseline 334a. The second distance 346 may be defined as the projected distance from the region of interest 326a to the baseline 334a, or a distance above or below the baseline in the mammogram example. Based on the first distance 344a, the algorithm 300 may determine a set of corresponding baseline 34b and first distance 344b in the right craniocaudal (RCC) view of FIG. 9B. Using the baseline 334b and first distance 344b determined in the second view of FIG. 9b, the algorithm 300 may further determine the corresponding band of interest 338 and display the band of interest 338 as shown. From within the band of interest 338 provided, the algorithm 300 may then enable a second selection or determination of the corresponding region of interest 326b in the second view. Using the region of interest 326b determined in the second view, the algorithm 300 may define a third distance 348 as the distance from the region of interest 326b to the baseline 334b, or the lateral distance from the nipple 330. Based on the first, second and third distances 344a-b, 346, 348, the algorithm 300 may be configured to determine the quadrant or the spatial coordinates of the region of interest 326a-b. Notably, while the respective baselines 334a-b, and/or the first distances 344a-b, of the first and second views of FIGS. 9A and 9B may be comparable in size and configuration, such parameters may be substantially different in other examples. In such cases, warping, or any other suitable process, may be used to reconfigure the respective volumes shown, as well as the respective parameters defined between commonly shared landmarks, to be in a more comparable form between the different views provided.

In a related modification, the algorithm 300 may be configured to superimpose a tracked region of interest 326 to a corresponding location on a reference structure 314 which may be a reference structure, prior sample structure, generic structure 324, or the like. As in previous embodiments, the algorithm 300 may initially determine control points (landmarks) that may be commonly shared by both the sample structure 312 and the reference structure 314. With respect to mammographic images 310, the control points may be defined as the nipple, the center of mass of the breast, the endpoints of the breast contour, or the like. Using such control points and a warping scheme, such as a thin-plate spline (TPS) modeling scheme, or the like, the algorithm 300 may be able to warp or fit the representations of the reference structure 314 to those of the sample structure 312. Once a region of interest 326 is determined and mapped within the sample structure 312, the spatial coordinates of the region of interest 326 may be similarly overlaid or mapped to the warped reference structure 314. Alternatively, a region of interest 326 that is determined within the reference structure 314 may be similarly mapped onto a sample structure 312 that has been warped to fit the reference structure 314.

The aforementioned concept of superimposing a tracked region of interest 326 to a corresponding location on a prior reference structure 314 may be extended to include multiple regions of interest. This enables one to readily determine the longitudinal progression in terms of growth or size and/or number of region(s) of interest over time. Initially there may be only one region of interest which may later grow or shrink. Mapping the initial image on the subsequent image enables accurate tracking of the growth or shrinkage of the region of interest. Additional regions of interest may develop over time and the algorithm 300 enables the user to accurately compare the region of interest longitudinally, i.e., over time. Importantly, the registration process may be automated to facilitate tracking the region of interest over time. However, even with a fully automated registration process it is desirable to prompt the user to manually confirm the registration or mapping of the region of interest and/or identified lesions within the region of interest. Alternatively, the fully automated system may allow the user to select the region of interest.

FIG. 16 shows a sample structure 312 with the region of interest shown in hashed lines. The user is able to access historical information regarding a region of interest using a pointing device such as a mouse, touch sensitive screen or the like. FIG. 17 is a table which was accessed by the user showing changes in the region of interest. In the example illustrated in FIG. 17 the comparison is made between the current measurements and the previous measurements for the region of interest.

The manual input from the user may consist of simply selecting a region of interest using a pointing device or a touch sensitive screen. In response to the manual input the algorithm 300 may display the outline or contours of a region of interest. In the event that the algorithm 300 cannot detect the outline of the region of interest it may prompt the user to manually trace the outline using a pointing device or the like or the algorithm 300 may simply place a circle or the like around the region of interest. The algorithm 300 uses the outline of the region of interest to automatically compute the size and/or volume of the region.

Alternatively, the algorithm 300 may use automated recognition techniques to identify and display one or more items of interest. The user is then prompted to accept or reject each item of interest. Alternatively, instead of prompting to accept or reject each item of interest, the user may be allowed to select a new or different region of interest.

Regardless of how the item of interest (region of interest) is identified (manual or automated) the user is then prompted to describe the item of interest using the knowledge representation corresponding to the imaging modality and/or the generic structure. To aid the user the knowledge representation presented to the user is both organ and modality specific. Alternatively, the knowledge representation presented to the user may be specific to organ or the imaging modality. Thus, terms which do not pertain to the imaging modality of the sample are not presented, nor are terms which do not pertain to the organs encompassing the region of interest. More specifically, the system may automatically select at least one diagnostic finding or prompt the user to select at least one diagnostic finding from the focused set knowledge representations. The system may retrievably store the at least one diagnostic finding in an electronic record such as an electronic medical record. The system may monitor (track) the electronic record for changes to the at least one diagnostic finding and/or the addition of new diagnostic findings, and may use such changes or new diagnostic findings to update the knowledge representation.

The selected generic structure may be related to the sample structure by imaging modality and one or more attributes such as size, dimensions, area, shape, volume, weight, density, location, anatomical organ, and orientation. The selected generic structure may have coordinate data defined therein.

Certain phenomena only occur in certain parts of certain objects such as anatomical organs. The knowledge of where certain phenomena are most likely to occur allows the system to provide a focused set of knowledge representations as a user interface (graphic or audio or both) to an image analyst. This focused knowledge representation allows the analyst to report the findings more efficiently. Focusing the knowledge representation may also be guided by other analytics from other data such as patient history, demographic, geolocation, genetic data etc. In addition to the structured knowledge representation in form of ontologies, the image analyst might add additional findings in form of free text entry. The image reporting system utilizes natural language analytics in form of statistical semantic analysis of text which is entered as free text and advise on patterns found in the free text. These patterns are basis for the evolution of the ontology. Further extensions of such mapping, marking and tracking may provide more intuitive three-dimensional representations of a sample structure 312, as shown for example in FIGS. 10A-10B. As a result of several iterations of mapping sets and subsets of known coordinate data to a sample structure 312, the algorithm 300 may be able to distinguish the different subcomponents of the sample structure 312 as separable segments, or subsets of data that are grouped according to like characteristics. For instance, in the sample structure 312 of FIGS. 10A-10B, each mammary gland may be defined as one segment 350. Such algorithms 300 may enable a user to navigate through three-dimensional layers of the sample structure 312 and select any point therein as a region of interest 326. In response, the algorithm 300 may determine the subcomponent or segment 350 located nearest to the region of interest 326 indicated by the user and highlight that segment 350 as a whole for further tracking, as shown for example in FIG. 10B.

Once at least one region of interest 326 has been determined and mapped, the algorithm 300 may further enable an annotation 352 of the region of interest 326 in an annotating step 354. For example, a physician viewing the two regions of interest 326 in FIGS. 9A-9B may want to annotate or identify the respective contents of regions of interest 326 as a cluster of microcalcifications and a spiculated nodule. Such annotations 352 may be received at the input device 206 in verbal form by way of a microphone, in typographical form by way of a keyboard, or the like. More specifically, the annotations 352 may be provided in the respective views of the sample structure 312 as plain text, graphics, playback links to audio and/or video clips, or the like. Once entered, each annotation 352 may be spatially associated and tracked with its respective region of interest 326 so as to be accessible and viewable in any related views depicting those regions of interest 326. Data associating each annotation 352 with its respective region of interest 326 may further be retrievably stored with the images 310 via an image server 106 and an image database 108 that is associated with, for example, a Picture Archiving and Communication System (PACS) in accordance with Digital Imaging and Communications in Medicine (DICOM). In an alternative embodiment, the algorithm 300 may be configured to receive an annotation 352 at the first instance of identifying a region of interest 326 and before any tracking of the region of interest 326 is performed to related views. Once the annotation 352 has been associated with the first determination of a region of interest 326, any corresponding regions of interest 326 tracked in subsequent views may automatically be linked with the same initial annotation 352. The algorithm 300 may also allow a user to edit previously established associations or relationships between annotations 352 and their respective regions of interest 326.

Turning back to the algorithm 300 of FIG. 3, step 356 of the algorithm 300 may configure an image reporting device 200 to allow generation of a report based on the mapped regions of interest 326 and accompanying annotations 352. As previously noted, the coordinate data of the generic structure 324 may conform to any common standard (taxonomy) for identifying spatial regions therein. For example, common standards for identifying regions of the breast may be illustrated by the coordinate maps of a generic breast structure in FIGS. 7A-7C. Once a sample structure 312 is mapped with such coordinate data, the algorithm 300 may be able to automatically identify the spatial location of any region of interest 326 or annotation 352 indicated within the sample structure 312. The algorithm 300 may then further expand upon such capabilities by automatically translating the spatial coordinates and/or corresponding volumetric data of the regions of interests 326 and the annotations 352 into character strings or phrases commonly used in a report.

With reference to FIG. 11A, an exemplary report 358 may be automatically provided in response to the regions of interest 326 and annotations 352 of FIGS. 9A-9B. As previously discussed, the mammographic representations of FIGS. 9A-9B depict two regions of interest 326 including a cluster of microcalcifications and a spiculated nodule. According to the coordinate system of FIGS. 7A-7C, the location of the cluster of microcalcifications may correspond to the superior aspect of the RLMO view at 11 o'clock, while the location of the spiculated nodule may correspond to the medial aspect of the LCC view at 10 o'clock. The algorithm 300 may use this information to automatically generate one or more natural language statements or other forms of descriptions indicating the findings to be included into the relevant fields 360 of the report 358, as shown in FIG. 11A. More specifically, the descriptions may include a location statement describing the spatial coordinates of the region of interest 326, a location statement describing the underlying object within the sample structure 312 that corresponds to the spatial coordinates of the region of interest 326, a descriptive statement describing the abnormality discovered within the region of interest 326, or any modification or combination thereof. The algorithm 300 may also provide standard report templates having additional fields 362 that may be automatically filled by the algorithm 300 (which may be manually over-ridden by the user) or manually filled by a user. For example, the fields 362 may be filled with data associated and stored for or with the patient and/or images, such as the exam type, clinical information, and the like, as well as any additional analytical findings, impressions, recommendations, and the like, input by the user while analyzing the images 310.

In further alternatives, the underlying object and/or abnormality may be automatically identified based on a preprogrammed or predetermined association between the spatial coordinates of the region of interest 326 and known characteristics of the sample structure 312 in question. The known characteristics may define the spatial regions and subregions of the sample structure 312, common terms (taxonomy) for identifying or classifying the regions and subregions of the sample structure 312, common abnormalities normally associated with the regions and subregions of the sample structure 312, and the like. Such characteristic information may be retrievably stored in, for example, an image database 108 or an associated network 102. Furthermore, subsequent or newfound characteristics may be stored within the database 108 so as to extend the knowledge of the database 108 and improve the accuracy of the algorithm 300 in identifying the regions, subregions, abnormalities, and the like. Based on such a knowledgebase of information, the algorithm 300 may be extended to automatically generate natural language statements or any other form of descriptions which preliminarily speculate on the type of abnormality that is believed to be in the vicinity of a marked region of interest 326. The algorithm 300 may further be extended to generate descriptions which respond to a user's identification of an abnormality so as to confirm or deny the identification based on the predetermined characteristics. For example, the algorithm 300 may indicate a possible error to the user if, according to its database 108, the abnormality identified by the user is not plausible in the marked region of interest 326. The algorithm 300 may use risk factors contained in the medical record of the patient as part of its decision criteria in indicating possible error or omission or to highlight potential concerns correlated with the risk factors. The user may choose to over-ride the error flag and may optionally provide a reason for over-riding the flag. Alternatively, the user may amend the identification of the abnormality. Thus, if the abnormality identified by the user is not commonly associated with a particular organ or with the patient's risk factors then the potential error will be flagged which may lead the user to revise the patient's risk factors. Moreover, the patient's risk factors indicate a high correlation or predisposition for a particular abnormality which was not identified by the user then the potential error will be flagged which may lead the user to more closely examine the region of interest for any over-looked abnormalities. One of the aspects of the present invention which should not be overlooked or minimized is the image reporting device and method of the present invention provides an image-based medical record which allows for tracking of diagnosis, decision on treatment and outcomes on a region of interest by region of interest (i.e. lesion by lesion) basis. The system may retrievably store the at least one diagnostic finding (diagnosis) in an electronic record. The system may monitor (track) the electronic record for changes to the at least one diagnostic finding and/or the addition of new diagnostic findings, and may use such changes or new diagnostic findings to update the knowledge representation. The system may monitor (track) the electronic record for changes to the patient outcome, and may use such changes to update the knowledge representation.

There are a variety of ways to access the stored information including selecting an (already identified) region of interest by, for example, touching the displayed region with a finger (touch sensitive screen) or using a pointing device. The image reporting device assigns each region of interest a unique label or identifier, and such identifier may also be used to access information pertaining to the diagnosis, treatment, and outcome of treatment. Once the user has selected a given region of interest, he/she is able to select prior annotations, display prior diagnosis, prior decisions on treatment and outcomes of such decisions—all on a region of interest by region of interest basis.

Access to prior annotations or the like may be made by, for example, a right mouse click or the like on the region of interest. Moreover, it should be noted that the image reporting system is intended to be used by both radiologists and oncologists. The radiologist uses the image reporting device 200 to enter diagnostic information and the oncologist uses the image reporting device to enter treatment information as well as treatment outcomes. In this manner the image reporting device facilitates collaboration and efficient sharing of information. In other alternatives, the algorithm 300 may automatically generate a web-based report 358, as shown in FIGS. 11B-11C for example, that may be transmitted to an image server 106 and/or an image database 108, and viewable via a web browser at a host 110, or the like. As in the report 358 of FIG. 11A, the web-based report 358 may be comprised of initially empty fields 362 which may be automatically filled by the image reporting system 200. The web-based report 358 may alternatively be printed and filled manually by a user. The report 358 may further provide an image representation of the sample structure 312 studied as a preview image 364. The report 358 may additionally offer other view types, as shown for example in FIG. 11C.

In contrast to the report 358 of FIG. 11B, the report 358 of FIG. 11C may provide a larger preview image 364 of the sample structure 312 and larger collapsible fields for easier viewing by a user. Providing such a web-based format of the report 358 may enable anyone with authorization to retrieve, view and/or edit the report 358 from any host 110 with access to the image source 216, for example, an image server 106 and an image database 108 of a Picturing Archiving and Communication System (PACS). In still further modifications, FIG. 12 schematically illustrates an image reporting system 400 that may incorporate aspects of the image reporting device 200, as well as the algorithm 300 associated therewith, and may be provided with additional features including integration with internal and/or external knowledge representation systems.

As shown, the image reporting system 400 may be implemented in, for example, the microprocessor 210 and/or memories 212-214 of the image reporting device 200. More specifically, the image reporting system 400 may be implemented as a set of subroutines that is performed concurrently and/or sequentially relative to, for example, one or more steps of the image reporting algorithm 300 of FIG. 3. An important aspect of the image reporting device 200 is that annotation, procedure history and outcomes are tracked on a lesion by lesion level. Selection of any lesion (by for example, right-clicking on a pointing device such as a mouse or the like) will enable the user to choose to display the previous diagnosis, treatment decisions, and longitudinal progression.

In this manner the user is able to see if the treatment regimen has been effective, where the current treatment regimen falls within the internal and/or external knowledge representation systems (ontology). In this manner the user will readily discern whether the current treatment is working and if not will see the next course of action recommended by the knowledge representation systems.

As shown in FIG. 12, once an image 401 of a sample structure that has been captured by an image capture device 104 is forwarded to the appropriate network 102 having an image server 106 and an image database 108, the image 401 may further be forwarded to the microprocessor 210 of the image reporting device 200. In accordance with the image reporting algorithm 300 of FIG. 3, a segmenting subroutine or segmenter 402 of the microprocessor 210 may process the image 401 received into subsets of data or segments 403 that are readily discernable by the algorithm 300. Based on the segmented image 403 of the sample structure and comparisons with a database 404 of generic structures 405, a mapping subroutine or mapper 406 may reconstruct a two- or three-dimensional image representation of the sample structure for display at the viewing device 208.

In addition to the image representation, the mapper 406 may also provide a semantic network 407 that may be used to aid in the general articulation of the sample structure, or the findings, diagnoses, natural language statements, annotations, or any other form of description associated therewith. For example, in association with an X-ray of a patient's breast or a mammogram, the semantic network 407 may suggest commonly accepted nomenclature for the different regions of the breast, common findings or disorders in breasts, and the like. The mapper 406 may also be configured to access more detailed information on the case at hand such that the semantic network 407 reflects knowledge representations that are more specific to the particular patient and the patient's medical history. For example, based on the patient's age, weight, lifestyle, medical history, and any other relevant attribute, the semantic network 407 may be able to advise on the likelihood whether a lesion is benign or requires a recall. Moreover, the semantic network 407 may display or suggest commonly used medical terminologies or knowledge representations that may relate to the particular patient and/or sample structure such that the user may characterize contents of the image representations in a more streamlined fashion.

Still referring to FIG. 12, the mapper 406 may refer to a knowledge representation broker or broker subroutine 408 which may suggest an appropriate set of terminologies (e.g. taxonomy), or knowledge representations (e.g. ontology), based on a structural triangulation or correlation of all of the data available. The broker subroutine 408 may access knowledge representations from external and/or internal knowledge representation databases and provide the right combination of knowledge representations with the right level of abstraction to the reader. More specifically, based on a specific selection, such as an anatomical object, made by the reader, the broker 408 may be configured to determine the combination of knowledge representation databases that is best suited as a reference for the mapper 406 and point the mapper 406 to only those databases. For a selection within a mammography scan, for instance, the broker subroutine 408 may selectively communicate with or refer the mapper 406 to one or more externally maintained sources, such as a Systematized Nomenclature of Medicine-Clinical Terms (SNOMED-CT) database 410, a Breast Imaging-Reporting and Data System (BI-RADS) database 412, a RadLex database 414 of common radiological terms, or any other external database of medical terminologies that may be used for characterizing findings within a sample structure and generating a natural language statement or any other form of description corresponding thereto. The mapper 406 may then refer to those knowledge representation databases in characterizing the selection for the reader using refined knowledge representations.

The broker 408 may also be configured to enable the reader to select one or more of the resulting knowledge representations to explore further refinements. The broker 408 may additionally be configured to determine an appropriate level of abstraction of the reader's selection based at least partially on certain contexts that may be relevant to the reader. The contexts may include data pertaining to the patient, the institution to which the reader belongs, the level of expertise of the reader, the anatomical objects in the immediate focus or view of the reader, and the like. The contexts may further include attributes pertaining to different interpretation styles and formats, such as iterative interactive reporting, collective reporting, and the like. Based on such contexts as well as the anatomical object selected by the reader, the image reporting system 400 may be able to provide more refined knowledge representations of the selected object that additionally suit the level of understanding or abstraction of the particular reader. The broker subroutine 408 may similarly access knowledge representations from an internally maintained dynamic knowledge representation database 416. The dynamic knowledge representation database 416 may further provide the broker 408 with the intelligence to provide the right combination of knowledge representations with the right level of abstraction.

Information generated by the mapper 406 may be provided in graphical form and, at least in part, as a transparent layer 418 such that the mapped information may be viewed at the viewing device 208 without obstructing the original image 401 upon which it may be overlaid. A user viewing the information displayed at the viewing device 208 may provide any additional information, such as regions of interest, annotations, statements of findings or diagnoses within the sample structure, and the like. Information input by the user, as well as any other data relevant to the patient, such as the patient's identification, demographic information, medical history, and the like, may be forwarded to a reporting subroutine or report engine 420 for report generation.

The report engine 420 may generate a report, for example, in accordance with the algorithm 300 disclosed in FIG. 3. Furthermore, the report engine 420 may forward the generated report to a medical record database 422 for storage and subsequent use by other care providers attending to the patient. As an additional or optional feature, the report engine 420 of FIG. 12 may be configured to forward a copy of the generated report to a tracking subroutine or case tracker 424.

Among other things, the case tracker 424 may serve as a quality tracking mechanism which monitors the amendments or findings in subsequent reports for any significant inconsistencies, such as mischaracterizations, oversights, new findings or diagnoses, disease progression or the like, and responds accordingly by adjusting one or more probability models associated with the particular knowledge representation in question. The case tracker 424 may monitor or track changes the electronic record such as changes to the at least one diagnostic finding and/or the addition of new diagnostic findings and/or treatment outcomes. The case tracker 424 may adjust the knowledge representation to reflect the changes to the diagnostic findings and/or the new diagnostic findings.

Probability models may be managed by the dynamic knowledge representation database 416 of the image reporting system 400 and configured to suggest knowledge representations that most suitably represents the anatomical object selected by the reader. Probability models may statistically derive the most appropriate knowledge representation based on prior correlations of data between selected elements or anatomical objects and their corresponding characterizations by physicians, doctors, and the like. Furthermore, the correlations of data and any analytics provided by the probability models may be dynamically updated, validated and invalidated according to any revisions as deemed necessary by the case tracker 424. For example, upon receipt of an alteration of the medical record, which reflects the performance of a treatment, the probability model of the knowledge representation may be validated or altered based on the content of the amendments of the medical record.

Based on the tracked results, the case tracker 424 may update the probability model within the dynamic knowledge representation database 416. For instance, a previous data entry of the dynamic knowledge representation database 416 which characterizes a structure with an incorrect statement or finding may be invalidated and replaced with a new data entry which correctly associates the structure with the new amendments or finding. Alternatively, the amendments or finding may be added to the existing statements as an additional finding for a particular combination of information. In such a manner, the case tracker 424 may continuously update and appropriately correct or enrich the representations stored in the dynamic knowledge representation database 416.

The case tracker 424 may use analytics to review free-form (natural language) text entered by the user such as diagnostic finding statements to study patterns of such analysis which may in turn be used to update the focused knowledge representation.

With such access to one or more of a plurality of knowledge databases 410, 412, 414, 416, the image reporting system 400 may be able to determine the best suited natural language statement or description for characterizing elements or findings within a sample structure. Moreover, the image reporting system 400 including at least, for example, a case tracker 424, a dynamic knowledge representation database 416 and a knowledge representation broker 408, may provide a feedback loop through which the image reporting algorithm 300 may generate reports with more streamlined terminologies, automatically expand upon its knowledge representations, as well as adjust for any inconsistencies between related reports and findings.

The medical record 422 may include a variety of patient information. The following list of patient information is intended to be representative but not exhaustive. The medical record may include some or all of the following: data corresponding to physical activities of the patient, patient genetic predisposition including DNA, medical history including prior cancer diagnosis, prior surgery, prior and current drug regimen, blood analysis information including pharmacological (drug absorption data), nutrition and the results of pathology reports. The term risk factors as used herein is intended to refer to one or more items of information from the medical record which either increase or decrease a person's predisposition to certain diseases. Such factors may include age, weight, family history, and the like. The data corresponding to physical activities may be collected using a Nike Fuel Band, Apple iWatch or like data collection devices such as known in the art.

Based on the aforementioned characterizing elements or findings within the sample structure the algorithm 300 and/or image reporting system may provide real time decision support by displaying recommendations based on guidelines for management of such findings. For example, in the context of the human lung, the Fleischner Society and the National Comprehensive Cancer Networks (NCCN) each provide guidelines for follow-up and management based on the size of the lesion and the presence of risk factors such as smoking, family history or the like. For each at least one region of interest, the system may automatically select follow-up care and/or prompt (allow) the user to select from a focused set of follow-up care options. The follow-up care is stored in the electronic record.

As will be explained below, the system monitors the electronic record for changes to the follow-up care and may use such changes to update the knowledge representation.

As shown in FIG. 14 the algorithm 300 prompts the user to select from one of the guidelines or enter a user specified instruction for follow-up care. In the illustration depicted in FIG. 14 the NCCN and Fleischner reflect two different guidelines which the user may select or enter a free-form (natural language) instruction in the box provided. The real time decision support may utilize guidelines found in a local database 426 (FIG. 12) or may access a third-party database 428 over the network. The follow-up care (selected or entered) is stored in the electronic record (electronic medical record).

In addition to showing the follow-up guidelines recommended by one or more third-party institutions such as Fleischner or NCCN, the algorithm 300 may provide a hyperlink to a knowledge base or the like providing additional insight into the guidelines. See, e.g. FIG. 15. The additional insight may, for example, include showing where the current treatment falls within an overall decision tree.

In still further modifications, one or more contents within the transparent layer 418 of the report may be configured to interact with a user through the user interface 204, or the like. For example, the transparent layer 418 may include an interactive knowledge representation displaying semantic relationships between key medical terminologies contained in statements of the report. Using a pointer device, or any other suitable input device 206, a user may select different terms within the report so as to expand upon the selected terms and explore other medical terminologies associated therewith. As the reader interacts with the knowledge representation, the broker might provide a different level of abstraction and a different combination of knowledge representations to assist in hypothesis building and provide information about probability of a malignancy to the reader.

A user viewing the report may also make new structural selections from within the image representation of the sample structure displayed. Based on the mapped locations of the user input, such selections made within the transparent layer 418 of the report may be communicated to the knowledge representation broker 408. More particularly, based on the new text selected by the user, the broker subroutine 408 may generate a new semantic network to be displayed within the transparent layer 418 of the report. Based on the new structure or substructure selected by the user, the broker subroutine 408 may determine any new set of medical terminologies, statements, findings, and the like, to include into the report.

The broker subroutine 408 may refer to any one or more of the knowledge representation databases 410, 412, 414, 416 shown in FIG. 12 in determining the ontologies and medical terminologies. Any required updates or changes to the report, or at least the transparent layer 418 thereof, may be communicated from the broker subroutine 408 to the report engine 420 such that a new and updated report is automatically generated for immediate viewing. Turning to FIGS. 13A-13B, another exemplary display or user interface that may be provided to the reader at the viewing device 208 is provided. More specifically, the display may follow a format that is similar to the display shown in FIGS. 4A-4B but with the additional feature of providing the reader with knowledge representations, for instance, in accordance with the image reporting system 400 of FIG. 12.

As in previous embodiments, a reader may choose to provide an annotation for a selected region of interest 326 by pointing to or indexing the region of interest 326 via the input device 206. In response to the anatomical object underlying or corresponding to the indexed region of interest 326, the image reporting system 400 of FIG. 12 may advise a focused set of knowledge representations most commonly associated with the anatomical object (e.g. an ontology). As shown in FIG. 13A, the knowledge representations may be presented to the reader in the form of a hierarchical menu or diagram showing semantic relationships, or the like. One or more of the knowledge representations displayed may be hierarchically configured and expandable to further reveal specific or more refined knowledge representations. For example, in the embodiment of FIG. 13A, the higher-level knowledge representation associated with the selected region of interest 326 may correspond to the lesion of a breast. Expanding upon this knowledge representation may then yield a plurality of common findings within the lesion of the breast. One or more of the resulting findings may also be expanded upon to reveal more refined subcategories, such as breast lumps, calcifications, nodules, sinuses, ulcerations, and the like. From the resulting subcategories, the reader may use the input device 206 to select the most appropriate finding that applies to the patient at hand. Once a knowledge representation is selected, the knowledge representation may be displayed as the annotation associated with the selected region of interest 326, as shown for example in FIG. 13B.

FIG. 18 is a flowchart of a method for integrating image analysis, longitudinal tracking of a region of interest and updating of a knowledge representation.

In step 1802, an image representation of a sample structure is retrieved from an electronic storage medium which such as an image database. The image database may be a PACS database (Picture Archiving and Communication System). In step 1804, the system automatically selects a generic structure from a database based on an imaging modality of the sample structure. At least one focused set of knowledge representations is stored in a second database. In some cases, the second database is the same database in which the generic structure is stored and, in some cases, the second database a different database. The knowledge representation is associated with or related to the selected generic structure by one or attributes such as imaging modality, size, dimensions, area, shape, volume, weight, density, location, anatomical organ, and orientation.

In step 1806, the selected generic structure is mapped by the system to the sample structure, and in step 1808 the system automatically determines at least one region of interest within the sample structure and/or allows the user to select a region of interest.

In step 1810 the system automatically selects at least one diagnostic finding and/or allows the user to select at least one diagnostic finding from the focused set knowledge representations. In other words, the system automatically selects at least one diagnostic finding. If the user disagrees with the automatically selected diagnostic finding, the user may select at least one diagnostic finding from a focused set of diagnostic findings. It should be understood that the diagnostic findings are focused to provide findings which are relevant in terms of imaging modality, anatomical organ or the like. If the user doesn't find the desired diagnostic finding in the focused set of findings then the user may enter a diagnostic finding using free-form text. In step 1812, the system retrievably stores the at least one diagnostic finding (the automatically selected diagnostic finding(s) or the diagnostic finding(s) selected or entered by the user) in the electronic record.

In step 1814, the system monitors or tracks the electronic record for changes to the at least one diagnostic finding and/or the addition of new diagnostic findings. The system uses the changes to the diagnostic findings and/or new diagnostic findings to update the knowledge representation in the second database.

The method may end at step 1814 or may optionally continue to step 1816 in which the coordinate data associated with the generic structure is used to generate natural language statements describing a location of the region of interest in the anatomy. The system automatically generates a diagnostic report based on the at least one diagnostic finding. The diagnostic report includes the natural language statements describing the location in the anatomy of the region of interest. The system stores the diagnostic report in the electronic record.

It should be understood that unless expressly stated otherwise, each of the method steps disclosed herein are performed by the system. Thus, the system automatically selects the region of interest, and the system monitors for changes to the electronic record.

In the aforementioned method of FIG. 18, the selected generic structure may be related to the sample structure by imaging modality and one or more attributes such as size, dimensions, area, shape, volume, weight, density, location, anatomical organ, and orientation. Moreover, the selected generic structure may have coordinate data defined therein.

In the method of FIG. 18, the knowledge representation may be specific to an anatomical organ in which the region of interest is located and/or the imaging modality used to capture the sample structure.

In the method of FIG. 18, the step of automatically selecting a generic structure (from among a plurality of generic structures) may be based on the imaging modality and/or a comparison of content of the sample structure to the content of the generic structure.

The method of FIG. 18 may end at step 1814 or step 1816 or the method may continue from either or both of these steps to step 1818 in which the system (algorithm) automatically selects follow-up care or allows the user to select from a focused set of follow-up care options for each at least one region of interest. More particularly, the user can change the automatically selected follow-up care option(s) automatically selected by the system by selecting at least one follow-up care option from a focused-sect of options or by entering a new follow-up care using free-form text. The system stores the follow-up care option(s) in the electronic record.

Step 1814 may also include checking for changes to the selected follow-up care and using such changes to update the knowledge representation in the second database. Additionally or alternatively, Step 1814 may include checking for changes to the previously stored treatment outcome and using such changes to update the knowledge representation in the second database.

FIG. 19 is a flowchart of another method for integrating image analysis, longitudinal tracking of a region of interest and updating of a knowledge representation. In step 1902 the system (algorithm) retrieve an image representation of sample structure depicting at least a portion of an anatomical organ from an image database. In step 1904 the system automatically determine at least one region of interest within the sample structure. If the user disagrees with the region(s) of interest automatically determined, the user is allowed the user to select a region of interest. Thereafter in step 1906, the system automatically selects at least one diagnostic finding. If the user disagrees, the user is allowed to select at least one diagnostic finding from a focused set of knowledge representations stored in a database or enter a diagnostic finding using free-form text. The focused set of knowledge representations is specific to one or more attributes of the sample structure such as imaging modality, size, dimensions, area, shape, volume, weight, density, location, anatomical organ, and orientation. In step 1908, the at least one diagnostic finding is stored in the electronic record. Then in step 1910, the system monitors or tracks the electronic record for changes and/or additions to the at least one diagnostic finding. If such changes are detected, the system uses the changes to update the knowledge representation.

Step 1910 or may optionally include checking for changes to treatment outcome and using such changes to update the knowledge representation in the database.

FIG. 20 is a flow diagram of a method for automatically improving a knowledge representation for an image reporting system.

In step 2002 the system records at least one diagnostic finding for a given region of interest in an electronic record. In step 2004, the system monitors or tracks the electronic record for changes to the at least one diagnostic finding for the region of interest. In step 2006, if such a change is detected the system automatically updates a knowledge representation stored in a database to reflect the changes to the at least one diagnostic finding. The method may terminate at step 2006 or may optionally include steps 2008-2014.

In step 2008 the system retrieves an image representation of sample structure depicting at least a portion of an anatomical organ from an image database. In step 2010 the system automatically determines at least one region of interest within the sample structure. Additionally or alternatively, the user is allowed to select a region of interest. In step 2012, the system automatically selects at least one diagnostic finding. Additionally or alternatively, the user is allowed to select at least one diagnostic finding from a focused set of knowledge representations specific to at least one of the anatomical organ and an imaging modality used to capture the image representation. Further still, the user may enter a diagnostic finding using free-form text. In step 2014, the system retrievably stores the at least one diagnostic finding in the electronic record.

Step 2004 may optionally include checking for changes to treatment outcome and using such changes to update the knowledge representation in the database.

Based on the foregoing, it can be seen that the disclosed method and apparatus provide an improved system and method for generating and managing image reports. The disclosed image reporting device and algorithms serve to automate several of the intermediary steps involved with the processes of generating and recalling image reports today. More specifically, the disclosed method and apparatus serves to integrate automated computer aided image mapping, recognition and reconstruction techniques with automated image reporting techniques. Furthermore, the disclosed method and apparatus aids in streamlining the language commonly used in image reporting as well as providing a means to automatically track subsequent and related cases for inconsistencies.

	Number	Date	Country
Parent	14093470	Nov 2013	US
Child	16363032		US

METHOD FOR INTEGRATING IMAGE ANALYSIS, LONGITUDINAL TRACKING OF A REGION OF INTEREST AND UPDATING OF A KNOWLEDGE REPRESENTATION

Information

Publication Number

Date Filed

Date Published

Inventors

Original Assignees

CPC

International Classifications

Abstract

Description

Claims

CROSS-REFERENCE TO RELATED APPLICATIONS

Continuations (1)