Patent Application
20240335178
This disclosure relates to detecting anomalies in a medical imaging system. The detection is based on an autoencoder that is trained to identify defects, malfunctions, and/or changes occurring in the medical imaging system.
The background description provided herein is for the purpose of generally presenting the context of the disclosure. Work of the presently named inventors, to the extent the work is described in this background section, as well as aspects of the description that may not otherwise qualify as prior art at the time of filing, are neither expressly nor impliedly admitted as prior art against the present disclosure.
Medical imaging systems, such as Positron Emission Tomography (PET) scanners, are widely used for diagnosis and clinical interventions. Before subjecting a patient to the hazard of medical radiation exposure with such a system, it is particularly important to ensure there are no defects, and to establish a known-good operating state. Scintillator-based, multichannel, gamma-ray detectors are typically used in a PET scanner. Due to the complexity in the structure and manufacturing process of such detectors, it is not uncommon for defects to occur that can impair the performance of the whole system. Only some of those defects are common and known, the rest may be novel and surprising.
Thus, it is conventional to check the detector performance daily by using the detectors to collect data from certain well-known radiation sources, and then produce a map of those detectors for visual inspection. More specifically, detector characteristics such as maps of counts per detector element, energy per detector, uniformity per detector region, etc. are constructed so as to be interpreted by a human or by an empirically designed algorithm. All of the known approaches require a priori knowledge of the meaning of the detector maps.
Furthermore, calibrations and corrections are often implemented through software computation based on data from a variety of sources and procedures to achieve uniform and stable performance. It is thus desirable to test the efficiency of those calibrations and corrections, and to spot an underperforming calibrating/correcting measure.
Therefore, methods and apparatus are desired to automatically learn and detect anomalies in a medical imaging system, including defects in various hardware and software modules, whether or not there is a priori knowledge of the defects.
The present disclosure relates to a method for detecting an anomaly related to a medical imaging device. The method comprises acquiring data from a plurality of detectors of the medical imaging device, applying the acquired data to a first autoencoder, and detecting, based on outputs from the first autoencoder, an anomaly related to the medical imaging device.
The disclosure additionally relates to an apparatus for detecting an anomaly related to a medical imaging device. The apparatus comprises processing circuitry. The processing circuitry is configured to acquire data from a plurality of detectors of the medical imaging device, apply the acquired data to an autoencoder, and detect, based on outputs from the autoencoder, an anomaly related to the medical imaging device.
Note that this summary section does not specify every embodiment and/or incrementally novel aspect of the present disclosure or claimed invention. Instead, the summary only provides a preliminary discussion of different embodiments and corresponding points of novelty. For additional details and/or possible perspectives of the invention and embodiments, the reader is directed to the Detailed Description section and corresponding figures of the present disclosure as further discussed below.
Various embodiments of this disclosure that are proposed as examples will be described in detail with reference to the following figures, wherein like numerals reference like elements, and wherein:
The following disclosure provides many different embodiments, or examples, for implementing different features of the provided subject matter. Specific examples of components and arrangements are described below to simplify the present disclosure. These are, of course, merely examples and are not intended to be limiting.
For example, the order of discussion of the different steps as described herein has been presented for clarity sake. In general, these steps can be performed in any suitable order. Additionally, although each of the different features, techniques, configurations, etc. herein may be discussed in different places of this disclosure, it is intended that each of the concepts can be executed independently of each other or in combination with each other. Accordingly, the present invention can be embodied and viewed in many different ways.
Furthermore, as used herein, the words “a,” “an” and the like generally carry a meaning of “one or more,” unless stated otherwise.
Aspects of this disclosure are directed to a method and apparatus for identifying anomalies based on unsupervised machine learning, which is achieved through an autoencoder, for example. The autoencoder is a particular type of artificial neural network that is configured to identify the salient characteristics of its input data. For example, used in facial recognition, it can identify the key measurements of a face image without being told beforehand what those measurements are. Used in medical imaging, it can identify key measurements of a lung image to estimate whether the image is after inhalation or exhalation, for example. One application of autoencoders is to analyze many datasets that should be similar, and then identify anomalies when one is not. This approach has been used in place of manual defect inspection in manufacturing applications, for example, using video of textile production, images of wind turbine performance data, and photographs of manufactured glass, or printed circuits.
In step 210, data (for example, PET detector maps) is acquired from detectors of a PET system during a well-known process, for example, a flood histogram of a point source or a lutetium background spectrum.
In step 220, an autoencoder is applied to the acquired data.
In step 230, outliers are identified based on the outputs of the autoencoder. The identified outliers might indicate “misbehaving” detectors.
During an autoencoder training phase, the autoencoder learns to identify the outliers. In step 240, a training dataset is generated from the data acquired from the detectors.
In step 250, the autoencoder learns based on the generated training dataset. When the training is completed, the learned autoencoder parameters are stored for use in the anomaly identification process.
The process of
In a non-limiting example, the autoencoder can learn a flood map in an X-ray imaging device and locate dead pixels. As another example, an energy map in a SPECT system can be learned by the autoencoder to identify the failure of various photosensors. In another non-limiting example, an autoencoder can be trained on a flood image from a 2D SPECT panel with a collimator attached, so as to check both the efficiency of the panel and the alignment of the collimator. By training and running an autoencoder on a CT sinogram obtained from a quality assurance phantom scan, the performance of a plurality of parts of the CT scanner can be checked, including the detectors, the rotation alignment, and the projection/sinogramming computation hardware.
Instead of “raw” data, the method 200 can be applied to “processed” data derived from various computation processes. Using an autoencoder trained with data generated by a software procedure, the performance quality of that procedure can be characterized. In a non-limiting example, an autoencoder can learn and operate with respect to a map of calibrated data, so that both hardware performance and calibration functionality can be examined. In another non-limiting example, after being trained with a map of calibrated data from a multichannel CZT detector, an autoencoder can check whether calibration parameters are appropriately selected, whether the calibration is operating as designed, etc. In the following, two preferable embodiments for examining the efficiency of a calibration will be described with reference to
Under different operating conditions, detectors can present significant individual variance in their characteristics. Normally, many kinds of calibration measures are taken to ensure a uniform output despite the varying operating conditions. For example, calibration is conducted to make the detectors work well under both high-count rate and low-count rate conditions. In the embodiments shown in
Then, the autoencoder can be applied to data divided into each time bin. In one embodiment, the outputs for a time bin can be compared against a reference standard to identify abnormal behavior (e.g., “drop-outs”) that occurs in time. The reference standard can be developed as a system calibration parameter, from a phantom scan or from background radiation, or as an average of many pervious patient scans, for example.
Alternatively, instead of using a reference standard, the outputs for many time bins can be compared to each other. In one example, the outputs from the autoencoder for each time bin are compared directly to identify abnormal behaviors. In another example, various processing such as clustering are applied to the outputs or the difference between the outputs for two time bins.
In one embodiment, the anomaly detection process shown in
The method shown in
If the answer at step 640 is “No,” the initial training dataset can be filtered to keep only high confidence normal data. In other words, only high confidence normal data will be used as a training dataset in a next iteration of training. Preferably, clustering is used in the selection of the training datasets. Further, a narrow cut can be used to keep only data well falling within a main cluster of the latent vectors or reconstruction error, for example, within ±0.5 standard deviations of the main cluster.
Steps 620-650 are repeated until stable outputs are derived from the autoencoder. At this point, the training of the autoencoder and the identification of anomalies are finalized. In a non-limiting example, the process shown in
In accordance with the methods described in this disclosure, either reconstructed data from the autoencoder or latent vectors calculated by the autoencoder are used to detect defects.
In one embodiment, to identify defects, the data reconstructed by the autoencoder is used to identify defects automatically, manually, or through another AE. In a non-limiting example, a pre-set threshold can be applied to automatically identify the image error. Firstly, the sum of squares difference (SSE) between the reconstructed data and the input can be calculated. Then, by comparing the calculated SSE with the threshold, non-uniform (i.e., defective) systems can be differentiated from uniform systems. As another non-limiting example, the data reconstructed from the autoencoder can be presented to an expert user (e.g., an operator or a service person who is responsible for reviewing quality-control images) to perform manual identification of defects. Alternatively, another autoencoder can be used to classify images as “normal” or “abnormal”. This autoencoder can be pre-trained using training datasets consisting of AE-reconstructed data.
In another embodiment, the latent vectors calculated by the autoencoder are directly used. Because the calculated latent vectors are highly compressed, computational analysis and comparisons can be faster and easier. For example, when specifications for the expected values of all or some of the latent vectors are set, a direct analysis of the compressed vector data can be conducted. As another example, clustering can be applied to the latent vectors, such as a k-means clustering algorithm. Alternatively, a power spectrum analysis or a principal-components analysis can be used.
As one example of system changes,
An autoencoder can be trained to overcome such an impairment. For each singles count rate map measured with an off-centered line source, a difference map compared to the singles count rate map measured with a well-centered line source can be produced. Using the difference maps as an input and the off-centered maps as a target, the autoencoder can learn to reconstruct the difference map.
Thus, when a measured count rate map is inputted to the trained autoencoder, the autoencoder can reconstruct a difference map. Thus, a corrected count rate map can be generated based on the reconstructed difference map and the measured count rate map.
On the basis of the corrected count rate map, defect detection can be carried out to identify the “real” anomalies in the system. The identification can be done by the same autoencoder, or by another autoencoder. Additionally or alternatively, the extent of the off-centeredness may be quantified from analysis of the reconstructed difference map.
Although the embodiments of this disclosure are described in the context of a whole system with a multitude of independent detectors, autoencoder-based defect detection can be applied to an individual detector, as part of a manufacturing process. By collecting plenty of data from a multitude of similar detectors to train an autoencoder, and operating the autoencoder on a single detector, performance of the detector can be evaluated.
Compared with the known approaches, autoencoder-based defect detection in accordance with the embodiments of the present disclosure uses machine learning to summarize the detector characteristics, without having any knowledge regarding the meaning of the acquired data, such as detector maps. An autoencoder takes over from a human or empirical algorithm the responsibility of determining what characteristics are most defining among normal and abnormal behaviors. Due to its automatic nature, the defect detection according to the embodiments of the present disclosure is more robustly automated. The autoencoder is able to detect unusual phenomena for which an empirical algorithm might not have been designed.
Each GRD can include a two-dimensional array of individual detector crystals, which absorb gamma radiation and emit scintillation photons. The scintillation photons can be detected by a two-dimensional array of photomultiplier tubes (PMTs) or silicon photomultipliers (SiPMs). A light guide can be disposed between the array of detector crystals and the photodetectors.
Each photodetector (e.g., PMT or SiPM) can produce an analog signal that indicates when scintillation events occur, and an energy of the gamma ray producing the detection event. Moreover, the photons emitted from one detector crystal can be detected by more than one photodetector, and, based on the analog signal produced at each photodetector, the detector crystal corresponding to the detection event can be determined using Anger logic and crystal decoding, for example.
In
The processor 870 can be configured to perform various steps of the methods described herein and variations thereof. The processor 870 can include a CPU that can be implemented as discrete logic gates, as an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other Complex Programmable Logic Device (CPLD). An FPGA or CPLD implementation may be coded in VHDL, Verilog, or any other hardware description language and the code may be stored in an electronic memory directly within the FPGA or CPLD, or as a separate electronic memory. Further, the memory may be non-volatile, such as ROM, EPROM, EEPROM or FLASH memory. The memory can also be volatile, such as static or dynamic RAM, and a processor, such as a microcontroller or microprocessor, may be provided to manage the electronic memory as well as the interaction between the FPGA or CPLD and the memory.
Alternatively, the CPU in the processor 870 can execute a computer program including a set of computer-readable instructions that perform various steps of the described methods, the program being stored in any of the above-described non-transitory electronic memories and/or a hard disk drive, CD, DVD, FLASH drive or any other known storage media. Further, the computer-readable instructions may be provided as a utility application, background daemon, or component of an operating system, or combination thereof, executing in conjunction with a processor, such as a Xeon processor from Intel of America or an Opteron processor from AMD of America and an operating system, such as Microsoft VISTA, UNIX, Solaris, LINUX, Apple, MAC-OS and other operating systems known to those skilled in the art. Further, CPU can be implemented as multiple processors cooperatively working in parallel to perform the instructions.
The memory 878 can be a hard disk drive, CD-ROM drive, DVD drive, FLASH drive, RAM, ROM or any other electronic storage known in the art.
The network controller 874, such as an Intel Ethernet PRO network interface card from Intel Corporation of America, can interface between the various parts of the PET scanner. Additionally, the network controller 874 can also interface with an external network. As can be appreciated, the external network can be a public network, such as the Internet, or a private network such as an LAN or WAN network, or any combination thereof and can also include PSTN or ISDN sub-networks. The external network can also be wired, such as an Ethernet network, or can be wireless such as a cellular network including EDGE, 3G and 4G wireless cellular systems. The wireless network can also be WiFi, Bluetooth, or any other wireless form of communication that is known.
Various techniques have been described as multiple discrete operations to assist in understanding the various embodiments. The order of description should not be construed as to imply that these operations are necessarily order dependent. Indeed, these operations need not be performed in the order of presentation. Operations described may be performed in a different order than the described embodiment. Various additional operations may be performed and/or described operations may be omitted in additional embodiments.
Embodiments of the present disclosure may also be as set forth in the following parentheticals.
Those skilled in the art will also understand that there can be many variations made to the operations of the techniques explained above while still achieving the same objectives of the invention. Such variations are intended to be covered by the scope of this disclosure. As such, the foregoing descriptions of embodiments of the invention are not intended to be limiting. Rather. any limitations to embodiments of the invention are presented in the following claims.
