Patent Application
20220245400
The invention generally relates to autonomous segmentation of three-dimensional nervous system structures from medical images of human anatomy, which is useful in particular for the field of computer-assisted surgery, surgical navigation, surgical planning, and medical diagnostics.
Image-guided or computer-assisted surgery is a surgical approach where the surgeon uses tracked surgical instruments in conjunction with preoperative or intraoperative images in order to indirectly guide the procedure. Image-guided surgery can utilize medical images acquired both preoperatively and intraoperatively, for example: from computer tomography (CT) or magnetic resonance imaging scanners.
Specialized computer systems can be used to process the medical images to develop three-dimensional (3D) models of the anatomy fragment subject to the surgery procedure. For this purpose, various machine learning technologies are being developed, such as a convolutional neural network (CNN) that is a class of deep, feed-forward artificial neural networks. CNNs use a variation of multilayer perceptrons designed to require minimal preprocessing.
A PCT patent application WO2017091833 (Arterys) discloses autonomous segmentation of anatomical structures, such as the human heart.
A US patent application US2016328630 (Samsung) discloses an object recognition apparatus and method that can determine an image feature vector of a first image by applying a convolution network to the first image.
In the field of image guided surgery, low quality images may make it difficult to adequately identify key anatomic landmarks, which may in turn lead to decreased accuracy and efficacy of the navigated tools and implants. Furthermore, low quality image datasets may be difficult to use in machine learning applications.
Computer tomography (CT) is a common method for generating a 3D volume of the anatomy. CT scanning works like other x-ray examinations. Very small, controlled amounts of x-ray radiation are passed through the body, and different tissues absorb radiation at different rates. With plain radiology, when special film is exposed to the absorbed x-rays, an image of the inside of the body is captured. With CT, the film is replaced by an array of detectors, which measure the x-ray profile.
The CT scanner contains a rotating gantry that has an x-ray tube mounted on one side and an arc-shaped detector mounted on the opposite side. An x-ray beam is emitted in a fan shape as the rotating frame spins the x-ray tube and detector around the patient. Each time the x-ray tube and detector make a 360° rotation and the x-ray passes through the patient's body, the image of a thin section is acquired. During each rotation, the detector records about 1,000 images (profiles) of the expanded x-ray beam. Each profile is then reconstructed by a dedicated computer into a 3D volume of the section that was scanned. The speed of gantry rotation, along with slice thickness, contributes to the accuracy/usefulness of the final image.
Commonly used intraoperative scanners have a variety of settings that allow for control of radiation dose. In certain scenarios high dose settings may be chosen to ensure adequate visualization of all the anatomical structures. The downside is increased radiation exposure to the patient. The effective doses from diagnostic CT procedures are typically estimated to be in the range of 1 to 10 mSv (millisieverts). This range is not much less than the lowest doses of 5 to 20 mSv estimated to have been received by survivors of the atomic bombs. These survivors, who are estimated to have experienced doses slightly larger than those encountered in CT, have demonstrated a small but increased radiation-related excess relative risk for cancer mortality.
The risk of developing cancer as a result of exposure to radiation depends on the part of the body exposed, the individual's age at exposure, the radiation dose, and the individual's gender. For the purpose of radiation protection, a conservative approach that is generally used is to assume that the risk for adverse health effects from cancer is proportional to the amount of radiation dose absorbed and that there is no amount of radiation that is completely without risk.
Low dose settings should be therefore selected for computer tomography scans whenever possible to minimize radiation exposure and associated risk of cancer development. However, low dose settings may have an impact on the quality of the final image available for the surgeon. This in turn can limit the value of the scan in diagnosis and treatment.
Magnetic resonance imaging (MRI) scanner forms a strong magnetic field around the area to be imaged. In most medical applications, protons (hydrogen atoms) in tissues containing water molecules create a signal that is processed to form an image of the body. First, energy from an oscillating magnetic field temporarily is applied to the patient at the appropriate resonance frequency. The excited hydrogen atoms emit a radio frequency signal, which is measured by a receiving coil. The radio signal may be made to encode position information by varying the main magnetic field using gradient coils. As these coils are rapidly switched on and off, they create the characteristic repetitive noise of an MRI scan. The contrast between different tissues is determined by the rate at which excited atoms return to the equilibrium state. Exogenous contrast agents may be given intravenously, orally, or intra-articularly.
The major components of an MRI scanner are: 1) the main magnet, which polarizes the sample, 2) the shim coils for correcting inhomogeneities in the main magnetic field, 3) the gradient system, which is used to localize the MR signal, and 4) the RF system, which excites the sample and detects the resulting NMR signal. The whole system is controlled by one or more computers.
The most common MR1 strengths are 0.3 T, 1.5 T and 3 T. The “T” stands for Tesla—the unit of measurement for the strength of the magnetic field. The higher the number, the stronger the magnet. The stronger the magnet the higher the image quality. For example, a 0.3 T magnet strength will result in lower quality imaging then a 1.5 T. Low quality images may pose a diagnostic challenge as it may be difficult to identify key anatomical structures or a pathologic process. Low quality images also make it difficult to use the data during computer assisted surgery. Thus, it is important to have the ability to deliver a high quality MR image for the physician.
There is a need to develop a system and a method for efficiently segmenting three-dimensional nervous system structures from intraoperative and presurgical medical images in an autonomous manner, i.e. without human intervention in the segmentation process.
One aspect of the invention is a method for autonomous segmentation of three-dimensional nervous system structures from raw medical images, the method comprising: receiving a 3D scan volume comprising a set of medical scan images of a region of the anatomy; autonomously processing the set of medical scan images to perform segmentation of a bony structure of the anatomy to obtain bony structure segmentation data; autonomously processing a subsection of the 3D scan volume as a 3D region of interest by combining the raw medical scan images and the bony structure segmentation data, wherein the 3D ROI contains a subvolume of the bony structure with a portion of surrounding tissues, including the nervous system structure; autonomously processing the ROI to determine the 3D shape, location, and size of the nervous system structures by means of a pre-trained convolutional neural network.
The method may further comprise 3D resizing of the ROI.
The method may further comprise visualizing the output including the segmented nervous system structures.
The method may further comprise detecting collision between an embodiment and/or trajectory of surgical instruments or implants and the segmented nervous system structures.
The nervous-system-structure segmentation CNN may be a fully convolutional neural network model with layer skip connections.
The nervous-system-structures segmentation CNN output may be improved by Select-Attend-Transfer gates.
The nervous-system-structures segmentation CNN output may be improved by Generative Adversarial Networks.
The received medical scan images may be collected from an intraoperative scanner.
The received medical scan images may be collected from a presurgical stationary scanner.
There is also disclosed a computer-implemented system, comprising: at least one non-transitory processor-readable storage medium that stores at least one processor-executable instruction or data; and at least one processor communicably coupled to at least one non-transitory processor-readable storage medium, wherein at least one processor is configured to perform the steps of the method as described herein.
These and other features, aspects and advantages of the invention will become better understood with reference to the following drawings, descriptions and claims.
Various embodiments are herein described, by way of example only, with reference to the accompanying drawings, wherein:
embodiment of the invention;
The following detailed description is of the best currently contemplated modes of carrying out the invention. The description is not to be taken in a limiting sense, but is made merely for the purpose of illustrating the general principles of the invention.
Several embodiments of the invention relate to processing three dimensional images of nervous system structures in the vicinity of bones, such as nerves of extremities (arms and legs), cervical, thoracic or lumbar plexus, spinal cord (protected by the spinal column), nerves of the peripheral nervous system, cranial nerves, and others. The invention will be presented below based on an example of a spine as a bone in the vicinity of (and at least partially protecting) the nervous system structures, but the method and system can be equally well used for nervous system structures and other bones.
Moreover, the invention may include, before segmentation, pre-processing of low quality images to improve their quality. This can be done by employing a method presented in a European patent application EP16195826 by the present applicant or any other pre-processing quality improvement method. The low quality images may be, for example, low dose computer tomography (LDCT) images or magnetic resonance images captured with a relatively low power scanner
The foregoing description will present examples related to computer tomography (CT) images, but a skilled person will realize how to adapt the embodiments to be applicable to other image types, such as magnetic resonance images.
The nerve structure identification method as presented herein comprises two main procedures in certain embodiments: 1) human-assisted (manual) training, and 2) computer autonomous segmentation.
The training procedure, as presented in
Next, the received images are processed in step 102 to perform autonomous segmentation of tissues, in order to determine separate areas corresponding to different parts of the bony structure, such as vertebral body 16, pedicles 15, transverse processes 14 and/or spinous process 11, as shown in
Then, in step 103, the information obtained from both original DICOM images and segmentation results is merged to obtain a combined image, comprising information about the tissue appearance and its classification (including assignment of structure parts to classes corresponding to different anatomy parts), for example in a form of a color-coded DICOM image 17, as shown in
Next, in step 104, from the set of slice images a 3D region of interest (R01) 18 is determined, that contains, for example, a volume of each vertebral level with a part of surrounding tissues including the nervous system structures and other structures such as muscles, vessels, ligaments, intervertebral discs, joints, cerebrospinal fluid, and others, as shown in
Then, in step 105, the 3D resizing of the determined ROI 18 is performed to achieve the same size of all ROI's stacked in the 3D matrices, each containing information about voxel distribution along X, Y and Z axes and the appearance and classification information data of bony structure, such as shown in the resizing (19A) of
Next, in step 106, a training database is prepared by a human, that comprises the previously determined ROIs and corresponding manually segmented nervous system structures.
Next, in step 107, the training database is augmented, for example with the use of a 3D generic geometrical transformation and resizing with dense 3D grid deformations. An example of such transformation for data augmentation 20 is shown in
Then, in step 108, a convolutional neural network (CNN) is trained with manually segmented images (by a human) to segment the nervous system structures. In certain embodiments, a network with a plurality of layers can be used, specifically a combination of convolutional with ReLU activation functions or any other non-linear or linear activation functions. For example, a network such as shown in
The segmentation procedure, as presented in
Next, in step 306, the nervous system structures are autonomously segmented by processing the resized ROI to determine the 3D size and shape of the nervous system structure(s), by means of the pretrained nervous-system-structure segmentation CNN 400, as shown in
In step 307 the information about the global coordinate system (ROI position in the DICOM dataset) and local ROI coordinate system (segmented nervous system structures size, shape and position inside the ROI) is recombined.
Next, in step 308, the output, including the segmented nervous system structures, is visualized.
Anatomical knowledge of position, size, and shape of nervous system structure(s) allow for real-time calculation of a possible collision detection with nervous system structure(s) (
One or more 3D ROI's can be presented to the input layer of the network to learn reasoning from the data.
The type of convolution layers 401 can be standard, dilated, or hybrids thereof, with ReLU, leaky ReLU or any other kind of activation function attached.
The type of upsampling or deconvolution layers 403 can also be standard, dilated, or hybrid thereof, with ReLU or leaky ReLU activation function attached.
The output layer 405 denotes the densely connected layer with one or more hidden layer and a softmax or sigmoid stage connected as the output.
The encoding-decoding flow is supplemented with additional skipping connections of layers with corresponding sizes (resolutions), which improves performance through information merging. It enables either the use of max-pooling indices from the corresponding encoder stage to downsample, or learning the deconvolution filters to upsample.
The general CNN architecture can be adapted to consider ROI's of different sizes. The number of layers and number of filters within a layer are also subject to change depending on the anatomical areas to be segmented.
The final layer for binary segmentation recognizes two classes: 1) nervous system structure, and 2) the background).
Additionally Select-Attend-Transfer (SAT) gates or Generative Adversarial Networks (GAN) can be used to increase the final quality of the segmentation. Introducing Select-Attend-Transfer gates to the encoder-decoder neural network results in focusing the network on the most important tissue features and their localization, simultaneously decreasing the memory consumption. Moreover, the Generative Adversarial Networks can be used to produce new artificial training examples.
The semantic segmentation is capable of recognizing multiple classes, each representing a part of the anatomy. For example the nervous system structure may include nerves of the upper and lower extremities, cervical, thoracic or lumbar plexus, the spinal cord, nerves of the peripheral nervous system (e.g., sciatic nerve, median nerve, brachial plexus), cranial nerves, and others.
The training starts at 501. At 502, batches of training 3D images (ROIs) are read from the training set, one batch at a time. For the segmentation, 3D images (ROIs) represent the input of the CNN, and the corresponding pre-segmented 3D images (ROIs), which were manually segmented by a human, represent its desired output.
At 503, the original 3D images (ROIs) can be augmented. Data augmentation is performed on these 3D images (ROIs) to make the training set more diverse. The input and output pair of three dimensional images (ROIs) is subjected to the same combination of transformations.
At 504, the original 3D images (ROIs) and the augmented 3D images (ROIs) are then passed through the layers of the CNN in a standard forward pass. The forward pass returns the results, which are then used to calculate at 505 the value of the loss function (i.e., the difference between the desired output and the output computed by the CNN). The difference can be expressed using a similarity metric (e.g., mean squared error, mean average error, categorical cross-entropy, or another metric).
At 506, weights are updated as per the specified optimizer and optimizer learning rate. The loss may be calculated using a per-pixel cross-entropy loss function and the Adam update rule.
The loss is also back-propagated through the network, and the gradients are computed. Based on the gradient values, the network weights are updated. The process, beginning with the 3D images (ROIs) batch read, is repeated continuously until an end of the training session is reached at 506.
Then, at 508, the performance metrics are calculated using a validation dataset—which is not explicitly used in training set. This is done in order to check at 509 whether not the model has improved. If it is not the case, the early stop counter is incremented by one at 514, as long as its value has not reached a predefined maximum number of epochs at 515. The training process continues until there is no further improvement obtained at 516. Then the model is saved at 510 for further use, and the early stop counter is reset at 511. As the final step in a session, learning rate scheduling can be applied. The session at which the rate is to be changed are predefined. Once one of the session numbers is reached at 512, the learning rate is set to one associated with this specific session number at 513.
Once the training process is complete, the network can be used for inference (i.e., utilizing a trained model for autonomous segmentation of new medical images).
After inference is invoked at 601, a set of scans (three dimensional images) are loaded at 602 and the segmentation CNN 400 and its weights are loaded at 603.
At 604, one batch of three dimensional images (ROIs) at a time is processed by the inference server.
At 605, the images are preprocessed (e.g., normalized, cropped, etc.) using the same parameters that were utilized during training. In at least some implementations, inference-time distortions are applied and the average inference result is taken on, for example, 10 distorted copies of each input 3D image (ROI). This feature creates inference results that are robust to small variations in brightness, contrast, orientation, etc.
At 606, a forward pass through the segmentation CNN 400 is computed.
At 606, the system may perform post-processing such as linear filtering (e.g., Gaussian filtering), or nonlinear filtering (e.g., median filtering, and morphological opening or closing).
At 608, if not all batches have been processed, a new batch is added to the processing pipeline until inference has been performed at all input 3D images (ROIs).
Finally, at 609, the inference results are saved and can be combined into a segmented 3D anatomical model. The model can be further converted to a polygonal mesh for the purpose of visualization. The volume and/or mesh representation parameters can be adjusted in terms of change of color, opacity, changing the mesh decimation depending on the needs of the operator.
The functionality described herein can be implemented in a computer-implemented system 900, such as shown in
The computer-implemented system 900, for example a machine-learning system, may include at least one non-transitory processor-readable storage medium 910 that stores at least one of processor-executable instructions 915 or data; and at least one processor 920 communicably coupled to the at least one non-transitory processor-readable storage medium 910. The at least one processor 920 may be configured to (by executing the instructions 915) to perform the steps of the method of
While the invention has been described with respect to a limited number of embodiments, it will be appreciated that many variations, modifications and other applications of the invention may be made. Therefore, the claimed invention as recited in the claims that follow is not limited to the embodiments described herein
| Number | Date | Country | Kind |
|---|---|---|---|
| 18205207.6 | Nov 2018 | EP | regional |
This application is a continuation of U.S. patent application Ser. No. 16/677,707, filed Nov. 8, 2019, entitled “Autonomous Segmentation of Three-Dimensional Nervous System Structures from Medical Images”, which claims benefit of European Application No. 18205207.6, filed Nov. 8, 2018, entitled “Autonomous Segmentation of Three-Dimensional Nervous System Structures from Medical Images”, the entire disclosure of which is incorporated herein by reference.
| Number | Date | Country | |
|---|---|---|---|
| Parent | 16677707 | Nov 2019 | US |
| Child | 17708907 | US |
