Medical image segmentation and atlas image selection

Abstract

Some embodiments relate to segmentation of medical images. For example, a medical image can be registered to a plurality of atlas images, after which segmentation functional components can be applied. A plurality of segmentations can be fused into a final overall segmentation. Atlas images can be selected based on high segmentation quality or low registration quality.

Description

Technical Field

[0001] The currently disclosed topics relate to methods for medical image segmentation, methods for selecting atlas images to be used in medical image segmentation methods, medical image segmentation systems, atlas image selection systems, and computer-readable media. Background Technology

[0002] Accurate and robust medical image segmentation remains a challenging task.

[0003] One approach to medical image segmentation is the multi-atlas-based image segmentation method. These methods can be constructed using a finite set of reference images for segmentation (called an atlas). Unfortunately, atlas-based methods often lack the required accuracy. Examples of registration methods and their applications in medical image segmentation are given in "Deformable medical image registration: a survey" by A. Sotiras et al. and "Symmetric diffeomorphic image registration with cross-correlation: evaluating automated labeling of elderly and neurodegenerative brain" by B. BAvants et al., both of which are included here by reference.

[0004] Another approach to medical image segmentation is through neural networks, such as convolutional neural networks. Examples of neural networks suitable for medical image segmentation are given in Groza et al.'s "Comparison of deep learning-based techniques for organ segmentation inabdominal CT images" and T. Brosch and A. Saalbach's "Foveal fully convolutionalnets for multi-organ segmentation," both of which are incorporated herein by reference.

[0005] Neural network-based methods can effectively construct accurate segmentation algorithms. Unfortunately, neural networks cannot generalize correctly to new datasets with slightly different properties. When a new image is presented to a trained neural network, even if the new image is slightly outside the range of images used to train the network, the resulting segmentation may be completely incorrect.

[0006] There is a need for a medical image segmentation method that allows for accurate segmentation on the one hand, and better generalizes to unfamiliar images on the other. Summary of the Invention

[0007] To address these and other issues, a medical image segmentation method using both a segmentation functional component and multiple atlas images is provided. Furthermore, a method for selecting the atlas images to be used in the medical image segmentation method is provided. Corresponding devices and software are also provided.

[0008] A machine learning-based segmentation algorithm was obtained, which produced accurate and robust segmentation of medical images and generalized well to new datasets with slightly different attributes. The improved generalization can also be used to reduce the amount of training data.

[0009] By registering images on atlases familiar to the segmentation functional unit (SFU), such as those where the SFU performs well, accurate segmentation is highly likely to be achieved. This avoids applying the SFU to images that would produce faulty segmentation patterns, such as those exhibiting a breakdown in segmentation capability. Registering images on multiple atlases reduces the probability that most atlas images will produce poor segmentation outputs and improves the accuracy of the process. For example, it prevents the random characteristics of the atlas or input images from dominating the results. Combining an ensemble of multiple predictors with a suitable voting scheme improves their predictive performance.

[0010] The advantage of the atlas image selection embodiment is that it can be performed after the segmentation functional unit has been trained. In fact, the training of the segmentation functional unit can be independent of atlas image selection. According to the embodiment, this makes it possible to enhance existing segmentation functional units.

[0011] For example, in one embodiment, the segmentation function may be a machine learning segmentation function. For example, training images may be used to train the segmentation function. For example, the segmentation function may be, for example, a neural network, such as a convolutional neural network (CNN). Instead of a neural network, the machine learning segmentation function may be a decision forest. The machine learning segmentation function may be an ensemble of neural networks. For example, the final result may consist of the results of the responses of the individual networks in the ensemble. For example, an ensemble of neural networks may be trained with different parameters.

[0012] A subset of n training images can be selected as an atlas. To segment a new image, it is registered to the selected n atlas images, and a segmentation function is applied to the registered images to obtain n segments. For example, these can be pixel or voxel segments. After applying the corresponding inverse transform to each of the n segments, the obtained segments can be fused, for example, using majority voting, to obtain the final segment.

[0013] The methods for segmentation and / or for selecting images from an atlas can be implemented on electronic devices, such as computers. The segmentation methods described herein can be applied in a wide range of practical applications, such as in clinical workstations for diagnosis, quantification, or treatment planning, or in network / cloud-based clinical applications.

[0014] Those skilled in the art will understand that this method can be applied to multidimensional image data (e.g., two-dimensional (2D), three-dimensional (3D), or four-dimensional (4D) images) acquired through various acquisition modalities, such as, but not limited to, standard X-ray imaging, computed tomography (CT), magnetic resonance imaging (MRI), ultrasound (US), positron emission tomography (PET), single-photon emission computed tomography (SPECT), and nuclear medicine (NM)).

[0015] Embodiments of this method can be implemented on a computer as a computer-implemented method, or in dedicated hardware, or a combination of both. Executable code for embodiments of the method can be stored on a computer program product. Examples of computer program products include memory devices, optical storage devices, integrated circuits, servers, online software, etc. Preferably, the computer program product includes non-transitory program code stored on a computer-readable medium for performing embodiments of the method when the program product is executed on a computer.

[0016] In one embodiment, the computer program includes computer program code adapted to perform all or part of the steps of an embodiment of the method when the computer program is run on a computer. Preferably, the computer program is embodied on a computer-readable medium. Attached Figure Description

[0017] Further details, aspects, and embodiments will be described by way of example with reference to the accompanying drawings. Elements in the drawings are shown for simplicity and clarity and are not necessarily drawn to scale. In the drawings, elements corresponding to those already described may have the same reference numerals. In the drawings,

[0018] Figure 1a An example of an embodiment of a medical image segmentation system is illustrated schematically.

[0019] Figure 1b An example of an embodiment of an atlas image selection system is illustrated schematically.

[0020] Figure 2a An example of an embodiment of a segmented functional component is illustrated schematically.

[0021] Figure 2b An example of an embodiment of image segmentation is illustrated schematically.

[0022] Figure 2c An example of an embodiment for determining segmentation quality is illustrated schematically.

[0023] Figure 3 An example of an embodiment of a method for medical image segmentation is illustrated schematically.

[0024] Figure 4a and Figure 4b An example of an embodiment of a method for selecting atlas images for use in a medical image segmentation method is illustrated schematically.

[0025] Figure 5a An example of an embodiment of a method for medical image segmentation is illustrated schematically.

[0026] Figure 5b An example of an embodiment of a method for selecting atlas images for use in a medical image segmentation method is illustrated schematically.

[0027] Figure 6a and Figure 6b An example of a medical image embodiment is illustrated schematically.

[0028] Figure 6c The Dice score of an example embodiment of a correctly functioning segmented functional component is illustrated schematically.

[0029] Figure 6d The Dice score of an example embodiment illustrating a segmented functional component that displays failure modes is schematically shown.

[0030] Figure 7a A computer-readable medium having a writable portion including a computer program, according to an embodiment, is illustrated schematically.

[0031] Figure 7b A representation of a processor system according to an embodiment is shown schematically.

[0032] Reference tag list:

[0033] 100 Machine Learning Systems

[0034] 110 Medical Image Segmentation System

[0035] 112 Segmented functional component storage device

[0036] 114 Image storage device

[0037] 130 processor system

[0038] 140 Memory

[0039] 150 communication interface

[0040] 160 Image Selection System

[0041] 164 Test Image Storage Device

[0042] 170 processor system

[0043] 180 memory

[0044] 190 Communication Interface

[0045] 210 Segmented functional components

[0046] 212, 213 Medical Input Images

[0047] Image segmentation 214, 215

[0048] 216 Real Image Segmentation

[0049] 220 comparator

[0050] 217 Segmentation Quality

[0051] More than 320 images

[0052] Images in the atlas 321-323

[0053] 312 Medical Images

[0054] 325 Registration Function Components

[0055] More than 330 registered images

[0056] Images registered 331-333

[0057] Image segmentation with over 340 registrations,

[0058] Image segmentation for registration (341-343)

[0059] Image segmentation with over 350 unregistration attempts.

[0060] 351-353 Image Segmentation for Unregistered Images

[0061] 361 Output Segmentation

[0062] More than 610 test images

[0063] 610 Test images that have not yet been selected

[0064] Test images 611-613

[0065] Segmentation of over 620 test images

[0066] 621-623 Test Image Segmentation

[0067] More than 630 segmentation qualities

[0068] Segmentation quality 631-633

[0069] 651 images in the image set

[0070] More than 640 registration qualities

[0071] Registration quality 641-643

[0072] Dice scores for the spine: 652, 654

[0073] Dice score for ribs: 662, 664

[0074] 400 A method for medical image segmentation

[0075] 410 Obtain a segmentation function unit, which is configured to receive a medical input image and generate image segmentation.

[0076] 420. Obtain multiple image sets.

[0077] 430 Receives medical images,

[0078] 440 The received image is registered to multiple atlas images, thereby obtaining multiple registered images and multiple corresponding registration transformations. The multiple corresponding registration transformations are configured to register the received image on multiple atlas images.

[0079] 450 The segmentation function is applied to multiple registered images to obtain image segmentation for multiple registrations.

[0080] 460. By applying the inverse of multiple registration transforms to multiple image segments, multiple image segments are obtained.

[0081] 470 Determine the output segmentation based on multiple image segmentations.

[0082] 500 A method for selecting atlas images to use in medical image segmentation methods,

[0083] 510 Obtain a segmentation function unit, which is configured to receive the input image and generate image segmentation.

[0084] 520. Obtain multiple test images and corresponding test image segments.

[0085] 530 The segmentation quality of multiple test images is determined by comparing the associated test image segmentation with the image segmentation generated by the segmentation function unit.

[0086] 540 Select one or more test images with segmentation quality exceeding the threshold from the test images as the atlas images.

[0087] 1000 computer-readable media

[0088] 1010 writable portion

[0089] 1020 Computer Program

[0090] 1110 (multiple) integrated circuits

[0091] 1120 Processing Unit

[0092] 1122 Memory

[0093] 1124 Application-Specific Integrated Circuit

[0094] 1126 Communication Components

[0095] 1130 Interconnection

[0096] 1140 Processor System Detailed Implementation

[0097] While the subject matter disclosed herein allows for many different forms of embodiments, one or more specific embodiments are shown in the accompanying drawings and will be described in detail herein. It should be understood that this disclosure is to be considered exemplary of the principles of the subject matter disclosed herein and is not intended to limit it to the specific embodiments shown and described.

[0098] In the following description, for ease of understanding, the elements of the embodiments are described in operation. However, it will be apparent that the corresponding elements are arranged to perform the functions described as being performed by them.

[0099] Furthermore, the subject matter disclosed herein is not limited to these embodiments, as features described herein or recited in mutually different dependent claims may be combined.

[0100] Figure 1aAn example embodiment of a medical image segmentation system 110 is illustrated schematically. For example, the image segmentation system 110 can be configured for image segmentation methods such as those shown herein. Figure 1b An example embodiment of an atlas image selection system 160 is illustrated schematically. For example, system 160 can be configured for a method of selecting atlas images as shown herein. For example, the atlas images selected in system 160 can be used for image segmentation in system 110.

[0101] The medical image segmentation system 110 and the atlas selection system 160 can be separate systems. The medical image segmentation system 110 and the atlas selection system 160 can be combined within the machine learning system 100. System 100 can be configured for atlas selection and image segmentation. The latter can be part of training segmentation functional units and / or segmenting new images.

[0102] For example, the medical image segmentation system 110 can be used in a production environment. For example, the medical image segmentation system 110 can be used by medical professionals or operators of medical imaging systems to generate segments of new images. The segmentation function in the medical image segmentation system 110 is typically a machine learning segmentation function, for example, which may be or has been trained on multiple training images. For convenience, it will be assumed that the segmentation function is of the machine learning type, but this is not strictly necessary. For example, according to embodiments, hand-designed segmentation functions, such as expert systems, may also be improved. The machine learning segmentation function may include a neural network, for example, configured to receive an image as input and produce a segmentation as output, but this is not necessary. For example, the segmentation function may include other machine learning functions (e.g., deep forest), or may include an ensemble of one or more machine learning functions. For example, when preparing the image segmentation system 110, an atlas image selection system 160 may be used. For example, system 160 may be used with a system for training the segmentation function of system 110. For example, system 160 may be further configured to train parameters of the segmentation function that can be used in the image segmentation system 110. However, the latter is optional; for example, segmentation feature training can be performed by different devices. In fact, it is advantageous to select atlas images after the segmentation feature training is complete. Interestingly, a fully trained segmentation feature can be obtained from any source (e.g., from a third party), and after the segmentation feature is fully trained, atlas images can be selected for it. In this way, the accuracy and robustness of any existing segmentation feature used for image segmentation can be increased later. This holds true even if the use of atlas images was not intended or anticipated during the design or training of the segmentation feature.

[0103] The image segmentation system 110 may include a processor system 130, a memory 140, and a communication interface 150. The image segmentation system 110 may be configured to communicate with a segmentation function component storage device 112 and an atlas image storage device 114.

[0104] The segmentation function component storage device 112 can be configured to store a segmentation function component that is configured to receive a medical input image and generate image segmentation. The atlas storage device 114 can be configured to store multiple atlas images.

[0105] Image selection system 160 may include processor system 170, memory 180, and communication interface 190. System 160 may include (e.g., in system 110) segmentation function component storage device 112 and test image storage device 164.

[0106] The segmentation function storage device 112 can be configured to store segmentation function components that are configured to receive medical input images and generate image segmentation.

[0107] The test image storage device 164 can be configured to store multiple test images and corresponding test image segments.

[0108] Storage devices 112, 114, and 164 can be local storage devices of system 110 or 160, such as local hard disk drives or memory. Storage devices 112, 114, and 164 can also be non-local storage devices, such as cloud storage devices. In the latter case, storage devices 112, 114, and 164 can be implemented as storage interfaces for non-local storage devices.

[0109] Systems 110 and / or 160 can communicate with each other via a computer network, and communicate with external storage devices, input devices, output devices, and / or one or more sensors. The computer network can be the Internet, an intranet, a LAN, a WLAN, etc. The system includes connection interfaces arranged to communicate as needed, either within or outside the system. For example, the connection interfaces can include connectors, such as wired connectors (e.g., Ethernet connectors, optical connectors, etc.) or wireless connectors (e.g., antennas, such as Wi-Fi, 4G, or 5G antennas).

[0110] The communication interface can be used to transmit or receive input images, atlas images, segmentation function component parameters, output images, etc. for segmentation.

[0111] The execution of systems 110 and 160 can be implemented in a processor system (e.g., one or more processor circuits, such as microprocessors, examples of which are shown herein). System 110 can be implemented in a single device that may or may not include a storage portion. System 160 can be implemented in a single device that may or may not include a storage device. Systems 110 and 160 can be implemented in a single system or device, etc. Systems 110 and / or 160 can include functional units that can be configured as elements (e.g., steps or portions, etc.) of embodiments of the image segmentation method and / or atlas selection method. The functional units can be implemented wholly or partially using computer instructions stored at systems 110 and 160 (e.g., computer instructions stored in the electronic memory of systems 110 and 160 and executable by the microprocessors of systems 110 and 160). In hybrid embodiments, the functional units are implemented partly in hardware (e.g., as coprocessors, such as segmentation functional component coprocessors, graphics coprocessors, etc.) and partly in software stored and executed on systems 110 and 160. The parameters and / or training data of the segmentation functional components can be stored locally at systems 110 and 160, or they can be stored in a cloud storage device.

[0112] Figure 2a Examples of embodiments of the segmentation function component 210 are illustrated schematically, such as those that may be used in embodiments of image segmentation methods and / or atlas image selection methods. For example, the segmentation function component 210 may be configured to receive a medical input image 212 and generate an image segmentation 214. For example, the segmentation function component may be configured to receive a 2D or 3D image I and generate annotations, such as segmentation A. The segmentation may be a binary image. For example, the image may be an image of a lung, and the segmentation may be a binary image indicating whether pixels / voxels of the input image correspond to a lung. The segmentation function component 210 may have corresponding annotations (A1, A2, ..., A...). m The image set (I1, I2, ..., I) m The segmentation functional components have been trained or are already trained. Training can be performed in a conventional manner. For example, a neural network can be trained (e.g., using backpropagation). Training can be performed using the Adam optimizer.

[0113] Figure 2b An example of an embodiment of image segmentation is illustrated schematically. Figure 2b The image shown is medical image 213. Medical image 213 shows objects of interest, such as the lungs, kidneys, and heart, but also objects of non-interest. In image 213, objects of interest are schematically indicated by squares. In this case, segmentation 215 is a binary image in which black pixels indicate the presence of objects of interest.

[0114] Instead of binary images, segmentation can also have two or more distinct categories, for example, to indicate more than one object of interest. For instance, a segment can be a two-dimensional or three-dimensional array indicating what object a pixel or voxel in the input belongs to. The size of the segment along its dimensions can be the same as the size of the input image, but it can also be smaller; for example, an output element can classify multiple input elements (e.g., pixels / voxels). The size of the segment along its dimensions can also be larger; for example, shape-based interpolation can be used to upsample the results.

[0115] In one embodiment, the input image may be an abdominal CT scan, and the objects of interest may be the liver, spleen, left kidney, and right kidney. The segmentation functional unit may be a neural network, such as a deep convolutional neural network (CNN) (e.g., U-shaped or F-shaped). The segmentation functional unit may also be trained on other objects and image modalities (e.g., arteries, bones, etc., or MRI, X-rays, etc.). The segmentation method may indicate common anatomical features in medical images, such as anatomical features present in anatomically normal (e.g., ordinary) individuals. The segmentation method may also, or alternatively, indicate medical abnormalities, such as tumors, fractures, etc.

[0116] Figure 3 An example of an embodiment of a method for medical image segmentation is illustrated schematically. Figure 5a An example of an embodiment of a method 400 for medical image segmentation is illustrated schematically. Figure 3 The data structures and data items that can be used in method 400 are shown. On the other hand, embodiments of method 400 can be used to manipulate... Figure 3 The data structures and data items shown are as follows.

[0117] Method 400 may include obtaining 410 segmentation function 210, which is configured to receive a medical input image and generate image segmentation.

[0118] For example, the segmentation function 210 can be a convolutional network or a ResNet-type architecture. Obtaining the segmentation function can include retrieving it from a storage device or receiving it from a computer network, etc. The segmentation function can be represented as a set of segmentation function parameters.

[0119] Typically, the segmentation function 210 is fully trained before atlas image selection, although this is not required. For example, after atlas selection, the segmentation function can be further trained, including input image registration. This could be fine-tuning, for example. In fact, during training, the segmentation function can be applied multiple times to multiple registered input images, and (e.g., as in one embodiment) the resulting multiple segments are fused, and an error signal is computed from the final fused segment. The latter error signal can be used for further training, for example, fine-tuning. One advantage of this approach is that the segmentation function is optimized for use with the selected atlas images. However, such fine-tuning is unnecessary; for example, atlas selection can be used as a way to improve the performance of the segmentation function without having to train it further.

[0120] Method 400 may include obtaining 420 or more atlas images 320. Figure 3 Images 321, 322, and 323 are shown. There may be more or fewer images in the atlas. For example, there may be 4 or more, 8 or more, 16 or more images in the atlas. The images in the atlas may have been selected according to the method used for selecting the images in the atlas. For example, the images in the atlas may have been selected for various positive attributes. For example, the images in the atlas may be well segmented. For example, the images in the atlas may represent good cross-sections of an image population.

[0121] Method 400 may include receiving 430 medical images 312. For example, medical images may be obtained from a medical imaging device (e.g., a CT scanner, an MRI scanner, an X-ray device, etc.). For example, the image may be represented as a 2D array of pixels or a 3D array of voxels. The image may be compressed or it may be in its original format, etc.

[0122] Method 400 may include registering the received image 312 440 to a plurality of atlas images to obtain a plurality of registered images 330 and a plurality of corresponding registration transformations, the plurality of corresponding registration transformations being configured to register the received image on the plurality of atlas images. Figure 3 Images 331, 332, and 333 showing the registration are illustrated. For clarity, in... Figure 3 The corresponding transformation is not shown in the diagram. Figure 3 A registration function component 325 is shown, configured to register image 312 on atlas image 320.

[0123] Image registration is the process of transforming a source image to better align it with a target image. In the case of functional unit 325, image 312 is the source image, and the atlas image serves as the target image. The registration functional unit can be configured to select permissible transformations from defined transformation categories. For example, the transformation categories can be, for instance, translation registration, rigid registration, similarity registration, affine registration, or non-rigid registration. Rigid transformations include translation and rotation. Similarity transformations include translation, rotation, and scaling. Registration can be selected by optimizing a loss function.

[0124] For the registration function component 325, elastic registration can be used. For example, the registration function component 325 can choose diffeomorphism. Elastic transformation allows for close alignment between the source and target images. Accordingly, it is desired that the segmentation quality of the atlas images can be matched by the transformed images.

[0125] After registration of a single input 312, multiple registered images are obtained, for example, one registered image for each atlas image. When registering image 312 to an atlas image, functional unit 325 generates the registered image and also generates a transformation mapping image 312 to the atlas image. For example, given input image I and atlas image I... i The transformation obtained can be selected from the transformation category, such that T i (I) at least to some extent related to image I i Alignment.

[0126] Method 400 may include applying segmentation function 210 450 to multiple registered images 330 to obtain multiple registered image segmentation 340. For example, segmentation function NN may be applied as NN(T) i (I)), where i exceeds the number of images in the atlas. The output of the segmentation function can be a 2D or 3D array of the same or smaller dimensions as the image 312. The elements of the array can indicate the segmentation determined by the segmentation function. Segmentation elements can be values ​​indicating the type of the corresponding pixel / voxel, but can also be vectors. For example, in the case where segmentation into p objects is required. The elements in the output array of the segmentation function can be p-dimensional vectors. The elements in the vectors can indicate the objects found. The sum of the vectors can be 1, or scaled to 1, etc.

[0127] Method 400 may include applying the inverse of multiple registration transforms 460 to multiple image segments 340 to obtain multiple image segments 350. For example, the method may calculate If the input and output sizes of the segmented functional unit are different, then the inverse transform is performed. It may be necessary to downsample to fit a smaller array. For example, the latter can be done through interpolation.

[0128] Method 400 may include determining an output segment 361 from multiple image segments 470. For example, this determination may include majority voting. For instance, image segment 350 may include a binary image; segment 316 may also include a binary image where pixel values ​​can be determined by the most frequently occurring values ​​of the corresponding pixels in segment 350. The same method can be used for multi-valued images instead of binary images. In the case where segment 350 is an array containing p-dimensional vectors (e.g., a 2D or 3D array), the corresponding vector elements can be averaged according to some averaging function. For example, given vector v... i The set, where each vector corresponds to the same element (e.g., pixel, etc.) in segment 350, can be used to determine the corresponding vector in segment 361 as 1 / n∑v i Interestingly, a strong function (majority function) can be approximated by using a power-mean (e.g., root mean square). For example, The power d is calculated component-wise. For example, d ≥ 2 can be chosen. The advantage of using higher values ​​of d (e.g., 2 or higher) is that it computes a majority-like determination while still allowing the use of vectors instead of single-value classifiers for multi-object segmentation.

[0129] As shown above, a segmentation function unit trained for image segmentation can be enhanced by registering images towards the atlas. The advantage is that the segmentation function unit is forced to provide multiple segmentations instead of a single segment. These segmentations are essentially for the same image, but this information is unavailable to the segmentation function unit.

[0130] Some types of segmentation functional units can be invariant to certain transformations. For example, if a segmentation functional unit includes a convolutional neural network, it can be invariant under translation, at least to some extent; however, this would not hold true for slightly more complex transformations. In fact, convolutional networks are generally not even invariant under rigid transformations, let alone elastic transformations (such as differential homeomorphisms). Therefore, if image 312 happens to be in a failure mode of segmentation functional unit 210, this is likely not true for the registered version of image 312, and even less likely for most registered images 330.

[0131] In one embodiment, the segmentation function is not invariant under the transform categories used to determine registration quality and / or for image segmentation. For example, in one embodiment, the segmentation function is invariant under a first transform set, and a second transform set is used to determine registration quality and / or for image segmentation, wherein the second set is larger than the first set, for example, the first set is a subset of the second set. For example, in one embodiment, the segmentation function includes a convolutional neural network, and the transforms used to determine registration quality and / or for image segmentation are larger than the transform set, for example, including rotation.

[0132] For example, assuming that image 312 has a probability of being missegmented of approximately 5%, and 10 images from the atlas are used, then the probability that most will be missegmented is approximately: The latter is a much smaller value. The above calculations assume independence between the applications of the segmentation functional units, but the primary cause of the dependency between multiple segments obtained from the segmentation functional units is imaging failure. For example, if the input image has very low quality (e.g., due to imaging equipment failure), then any application for any registered segmentation functional unit is likely to fail. However, this failure mode can be attributed to reasons different from those of the segmentation functional techniques. The number of ten atlas images above is an example. A reduction in the probability of failure can also be achieved with fewer atlas images.

[0133] Image sets can be selected from the images used to train the segmentation functional unit. For example, image sets can be selected randomly. The advantage of selecting image sets from training images is that the segmentation functional unit is likely familiar with those images. However, better results can be obtained by using more careful selection, for example, as described in this paper.

[0134] Figure 4a An example of an embodiment of a method for selecting atlas images to be used in a medical image segmentation method is illustrated schematically. Figure 5b An example embodiment of a method 500 for selecting atlas images for use in a medical image segmentation method is illustrated schematically. Atlas images can be selected for methods such as method 400.

[0135] For example, method 500 can use Figure 4a The data items and structures shown are, and vice versa. Method 500 can also be used as described below. Figure 4b The data shown.

[0136] Method 500 may include obtaining a segmentation function 210, which is configured to receive an input image and generate image segmentation. The segmentation function 210 may be the same as the segmentation function to be used in production. As described above, the segmentation function may include a neural network, such as U-net or F-net, CNN, etc.

[0137] Method 500 may include obtaining 520 or more test images 610 and corresponding test image segments 620. For example, the multiple test images 610 may be obtained from the training of the segmentation function 210. For example, the segmentation function 210 may be trained on multiple pairs of training images and training image segments. For example, the backpropagation method may be used to train a neural network on a training set. Typically, for testing purposes, such as testing the convergence of the segmentation function, some image sets are set aside. Test images 610 may be obtained from images used for training and / or from images used to test the segmentation function. Preferably, the test images 610 are obtained from the same or similar image distribution on which images can be drawn for training the segmentation function.

[0138] Test images 610 can be all images used for training and / or testing of the segmentation functional unit 210, especially if these images are relatively few, for example, less than about 100. Test images 610 can also be a subset, for example, a random selection of images used for training and / or testing. The latter is useful if the number of training / test images is large.

[0139] Method 500 may include determining the segmentation quality 630 of 530 test images by comparing an associated test image segmentation 620 with an image segmentation generated by the segmentation function. For example, test image 610 may have an associated segmentation, such as ground truth segmentation 620. The latter may also have been used in training or testing. There are several ways to calculate segmentation quality. One way to do this is to calculate a Dice score between the ground truth segmentation and the segmentation generated by the segmentation function. For example, the segmentation quality of image 611 can be obtained by calculating the Dice score between segmentation 621 and segmentation 631. A high Dice score indicates good overlap, thus indicating high segmentation quality.

[0140] Method 500 may include selecting one or more test images from the 540 test images that have segmentation quality exceeding a threshold as an atlas image. For example, a Dice score exceeding a threshold, such as a Dice score exceeding 0.8, may be required. The threshold may be predetermined. The threshold may be dynamic. For example, the atlas images may be selected from the best-performing images. For example, a certain proportion of the worst-scoring images may be discarded.

[0141] The inventors discovered that, in practice, most images perform well in segmentation, but a certain percentage (e.g., about 5%) of the images show worse segmentation than the majority. By selecting an atlas of images from the best-performing 95%, 90%, or 80% of the images, the input image 312 is prevented from being registered on images where the segmentation function performs poorly, thus making it highly likely that the segmentation function will also perform poorly on the registered images.

[0142] For example, in one embodiment, the worst-performing portion of image 610, such as the bottom 5%, is discarded. Images from the atlas can then be selected from the remaining images; this can be done randomly, for example, by randomly selecting 10 images.

[0143] For example, in addition to segmentation quality, or in place of segmentation quality, additional criteria can be imposed on the selected atlas images.

[0144] Figure 2c An example of an embodiment for determining segmentation quality is illustrated schematically. An image 212 with an associated ground truth image segmentation (e.g., true segmentation) is shown. Image 212 is segmented by segmentation function 210 to obtain a generated image segmentation 214. The generated image segmentation 214 can then be compared with the ground truth image segmentation 216 (e.g., using comparator 220) to obtain segmentation quality 217. For example, comparator 220 can calculate a Dice score.

[0145] Figure 4b An example of an embodiment of a method for selecting atlas images for use in medical image segmentation is illustrated schematically. Figure 4b Image 610' is shown that has not yet been selected as an atlas image and has not yet been discarded for other reasons (e.g., due to poor segmentation quality). Figure 4b The image 651, which has been selected as an image in the atlas, is also shown. For example, image 651 may have been selected because it has the best segmentation quality, or it may have been randomly selected from images with good segmentation quality.

[0146] The registration function is applied to image 610' on the selected image 651. Registration quality is obtained by comparing the registered test image with the selected test image 651. In this way, registration quality 640 is obtained. The comparison can use a comparator 220, such as a Dice score. Other similarity measures can be used instead of the Dice score, such as correlation.

[0147] Interestingly, and unlike segmentation quality, atlas images with low registration quality can be selected. Low registration quality indicates that the image is quite different from the images already selected as atlas images. For example, the image might have the lowest registration score. For instance, images with the best registration quality can be discarded, and then one or more random images can be selected. For example, images can be selected from 50% of the images with the lowest registration quality.

[0148] In one embodiment, the registration performed in method 400 (e.g., in registration function 325) can be selected from the same transformation category as that performed for calculating registration quality. Specifically, the registration type for atlas image selection and for image segmentation can be translation-only, such as shift, rigid, similarity, affine, or non-rigid (e.g., elastic, such as differential homeomorphism). The latter transformation can be non-elastic or at least much less elastic.

[0149] Interestingly, for the registration performed in method 400 (e.g., in registration function 325), it is possible to select registration from a different category of transformation (e.g., a larger category of transformation) than the registration performed to calculate the registration quality. Specifically, the former could be elastic registration, such as a differential homeomorphism. The latter could be inelastic or at least much less elastic. For example, rigid registration could be used to calculate the registration quality. Rigid transformations only allow transformations and rotations. Slightly larger categories could be allowed, such as transformations, rotations, and scaling, or affine transformations.

[0150] The elastic registration in method 400 allows images to match the atlas images as closely as possible, thereby achieving segmentation quality comparable to that of the atlas images. On the other hand, rigid transformation allows for the discovery of a wider variety of images, since elastic registration would make the images look too similar.

[0151] exist Figure 4b In this process, multiple test images are compared with a single atlas image. For example, after a first atlas image has been selected based solely on segmentation quality, this can be used to select a second atlas image. This can be used, for example, to add atlas images that differ from previously selected specific atlas images in terms of registration. Subsequent atlas images can be randomly selected from the already selected atlas images, or, for example, each selected image can be used iteratively, and so on.

[0152] Another approach is to compare the test image 610' with multiple atlas images, or even with all atlas images selected so far, instead of comparing it with a single selected atlas image 651. For example, the registration quality between the test image 610' and the selected atlas images can be calculated. The overall registration quality of the test image 610' can then be determined. This could be, for example, the average registration quality. A high average indicates generally high registration quality. Power arithmetic can be used instead of the arithmetic mean, highlighting atlas images that are very different from some images, even if they are similar to most. Selecting another atlas image with low overall registration quality is more likely to result in independent segmentation by the segmentation function. On the other hand, if that other atlas image has high segmentation quality, the segmentation function is more likely to perform well in it.

[0153] Selection based on global registration can be achieved by discarding overly similar images and randomly selecting from the remaining images. Other selection methods are also possible, such as those described in this paper.

[0154] For example, the following steps can be used

[0155] 1. Select a random pool of test images, for example, 100 or 1000 training images with associated real segmentations.

[0156] 2. Discard the 'a%' of the pool with the worst split quality. For example, 'a' could be 5%.

[0157] 3. Select random images from the current pool and remove the selected images from the pool.

[0158] 4. Calculate the overall registration quality of the test image pool compared to the selected atlas images.

[0159] 5. Discard images with excessively high overall registration quality from the pool.

[0160] 6. Unless the pool is empty or a sufficient number of atlas images have been selected, proceed to section 3.

[0161] Step 2 can use absolute scores instead of relative scores. For example, images with Dice scores below 0.8 between the generated and true segments can be discarded.

[0162] Step 4 can be performed efficiently because the registration quality between each test image and the atlas image can be stored, and it does not need to be calculated again. Therefore, if a single atlas image is added, only the registration quality between the test image and the added atlas image needs to be calculated.

[0163] Step 5 can remove b% of the pool (i.e., 5% or higher, such as 25%). However, an absolute threshold can also be used instead of a fixed percentage. The absolute threshold can be determined empirically.

[0164] Instead of random selection, for example in step 3, an image with the best segmentation quality and / or the worst registration quality, or a combination thereof (e.g., summation, etc.), can also be selected. The atlas image selection algorithm above is an example, but different embodiments are possible.

[0165] Generally, the boundaries of segmentation quality and / or registration quality, whether absolute or relative, can depend on the difficulty of the segmentation problem, the quality of the segmentation functional unit (e.g., the amount of training data), the acceptance of false positives, and so on. If the segmentation functional unit includes a neural network, its quality can depend on its depth. Boundaries can be empirically established by splitting the training image set as test data and evaluating the resulting segmentation system. In various embodiments of systems 110 and 160, the communication interface can be selected from a variety of alternatives. For example, the interface could be a network interface to a local area network or wide area network (e.g., the Internet), a storage interface to an internal or external data storage device, a keyboard, an application programming interface (API), etc.

[0166] Systems 110 and 160 may have a user interface, which may include well-known components such as one or more buttons, a keyboard, a display, a touchscreen, etc. The user interface may be configured to adapt to user interaction for configuring the system, training a segmentation function on a training set, applying the system to new image data, or selecting images from an atlas, etc.

[0167] The storage device can be implemented as an electronic memory (such as flash memory) or a magnetic memory (such as a hard disk). The storage device may include multiple discrete memories that together constitute storage devices 140 and 180. The storage device may include temporary memory, such as RAM. The storage device may be a cloud storage device.

[0168] System 110 can be implemented in a single device. System 160 can be implemented in a single device. Typically, both systems 110 and 160 include a microprocessor that executes appropriate software stored at the system; for example, this software may have been downloaded and / or stored in a corresponding memory, such as volatile memory like RAM or non-volatile memory like flash memory. Alternatively, the system can be implemented wholly or partially in programmable logic, for example, as a field-programmable gate array (FPGA). The system can be implemented wholly or partially as a so-called application-specific integrated circuit (ASIC), for example, an integrated circuit (IC) customized for its specific purpose. For example, the circuit can be implemented using CMOS, for example, using a hardware description language such as Verilog, VHDL, etc. Specifically, systems 110 and 160 may include circuitry for evaluating partitioned functional components.

[0169] The processor circuitry can be implemented in a distributed manner, for example, as multiple sub-processor circuits. The storage device can be distributed across multiple distributed sub-storage devices. Part or all of the memory can be electronic memory, magnetic memory, etc. For example, the storage device can have volatile and non-volatile components. A portion of the storage device can be read-only.

[0170] In one embodiment, test-time data augmentation is included, for example, in method 400. The segmentation feature (e.g., segmentation feature 210) can be trained using conventional segmentation features; for example, the segmentation feature can be configured for conventional machine learning algorithms. For example, the machine learning algorithm could be a decision forest. For example, the machine learning algorithm could be a neural network. For example, the segmentation feature could include a neural network, such as a convolutional neural network (CNN) (like U-net or F-net, etc.). Atlas images can be selected from training images. Several further optional refinements, details, and embodiments are shown below.

[0171] In one embodiment, n ≥ 2 atlases of images are used, such as n = 10, etc. A new image for segmentation can be registered to the n atlases, resulting in n invertible transformations. A segmentation function (e.g., a CNN) is applied to each of the n transformed images, resulting in n annotated images, such as segmentation images. The inverse transformation generated by the registration is applied to the annotated images, and the labels of the transformed annotated images are fused to obtain the final segmentation.

[0172] This basic idea can also be applied more generally to other tasks, such as image classification (e.g., using a ResNet-type neural network architecture). In this case, the inverse transformation step after applying the segmentation functional unit can be skipped. The segmentation functional unit can be trained using corresponding annotations (A1, ..., A...). mA set of m 2D or 3D images (I1, ..., I...) m Training of 2D / 3D segmentation functional units for image segmentation can be performed in a conventional manner. For example, a neural network can be trained (e.g., using backpropagation, such as using the Adams optimizer). m images and their corresponding annotations can be used together for segmentation functional unit training, or they can be subdivided into training and test sets.

[0173] Atlas selection can choose a subset n (≤m) of m images as an atlas. Various criteria can be used to select the individual images used as the atlas. One criterion could be the segmentation quality observed during training (e.g., the Dice score of images trained as part of the training data). Another criterion could be the segmentation quality observed during testing (e.g., the Dice score of images trained as part of the test data). Yet another criterion could be the image differences. For example, rigid registration can be applied to the training images (e.g., to the atlas candidates), and then candidates with the minimum mutual annotation overlap in Dice score after registration to the already selected candidates can be selected. This process can be iteratively repeated to select n atlas images.

[0174] Numerous algorithms have been proposed and applied to medical image registration and image transformation for image-to-atlas conversion. Recently, registration algorithms utilizing deep learning and neural networks have also been employed. Algorithms that generate differential homeomorphic transformations have the advantage of easily obtaining the inverse transformation. Depending on the images being processed, similarity measures applicable to registration of images with the same modality (e.g., cross-correlation) or different types of images (e.g., mutual information) can be used. Applying (non-rigid) registration algorithms to image I... new and atlas images (I1, ..., I n This generates n transformations (T1, ..., T2) that transform the new image into the space of the corresponding atlas. n For example, as T i (I new ).

[0175] When an atlas image has a larger field of view than the new image to be segmented, the atlas image can be used to supplement the field of view of the new image.

[0176] In addition to spatial image transformation, the new image I new The grayscale values ​​of the image can also be transformed to better represent the atlas image. This transformation can be included in transformation T. i For example, a parameter transformation can be applied to the intensity; the parameters can be determined such that the histogram of the new image obtained after the intensity transformation corresponds to the atlas image.

[0177] The transformed image (T1(I) can be segmented using a previously trained segmentation function. new ), ..., T n (I new (e.g., multiple registered images), thereby generating annotations (A) 1，new A n，new After segmentation, the annotations can be transformed back into the space of the new image, thus generating annotations. The segmentation functional components can be implemented using, for example, neural networks (e.g., U-net or F-net architectures), but are not limited to any particular type of network. Alternatively, other neural networks can be used.

[0178] Tag fusion: given annotation The final annotation of the new image can be constructed (e.g., by majority voting); for example, a label corresponding to the most frequent label at position x in the annotation to be fused can be assigned to each voxel at position x. Alternatively, other label fusion techniques can be used. In one embodiment, information about the fusion (e.g., voting) can be used for further processing or displayed to a user. For example, a second label or the top k labels can be displayed so that the operator knows what the possible second option is. The latter can be used for quick feedback. For example, when the operator encounters an incorrect segmentation, but it is highly likely that the second option or the top k options are correct, this information can be provided (e.g., using the input device of the segmentation device). For example, the operator can use a mouse, etc., to indicate that a specific object or pixel / voxel belongs to the second most frequent label, etc. This information can later be used to retrain or fine-tune the segmentation functionalities.

[0179] An example of rib and spine segmentation in 3D CT images is used to illustrate the implementation; Figure 6a A visual representation of this CT scan is shown. Specifically, the F-net neural network has been trained to handle this task. Segmentation is performed using a segmentation functional unit that includes the neural network. Figure 6a The image was obtained, and the Dice score between the segmentation function component output and the expert segmentation was calculated. Figure 6c The diagram shows the Dice score for the spine at 652 and the Dice score for the ribs at 662. Note that both scores are high, indicating that the segmented functional components are... Figure 6a It performs well on images.

[0180] Data augmentation was used during the training of the neural network in the segmentation functional unit; in this case, the CT dataset was rotated by a maximum of 7°. Within this range, the segmentation functional unit is expected to produce accurate output. Figure 6b It shows the relationship with Figure 6aThe same image, but with a significantly different orientation; here it's rotated 45°. Note that apart from this rotation, the image is identical. When the segmentation function is applied... Figure 6b At that time, the Dice score obtained was much smaller. Figure 6d The diagram shows the Dice score for the spine at 654 and the Dice score for the ribs at 664. Note that both scores are low, indicating that the segmented functional components are in a low position. Figure 6b The image performs poorly. Note that the score for the spine is approximately half of the previous value, indicating a breakdown in predictive power for spine objects in this case.

[0181] Even if Figure 6b rigid registration to similar Figure 6a The atlas images also return the segmentation capabilities of the segmentation functional unit. More subtle prediction breakdowns (e.g., caused by unusual anatomical variations) can be addressed using non-rigid registration. Applying the segmentation functional unit to a transformed version of the label fusion in the test images improved the segmentation results.

[0182] For example, the segmentation function, atlas images, or test images can be obtained from an electronic storage system, which can be internal or external, such as accessible via computer cable or computer network. The segmentation function can be trained on images (e.g., medical images obtained from medical imaging equipment, such as a CT scanner). Receiving the medical images for segmentation can be performed in the same manner as described above. This reception can be via an API or other interface (which can be internal or external). Images can be received from imaging equipment in a medical terminal. Image registration, determining registration or segmentation quality, etc., can be performed on electronic devices such as computers. During training and / or application, the neural network used for the segmentation function can have multiple layers (which may include, for example, convolutional layers). For example, the neural network in the segmentation function can have at least 2, 5, 10, 15, 20, or 40 hidden layers or more, etc. The number of neurons in the neural network can be at least, for example, 10, 100, 1000, 10000, 100000, 100000 or more, etc.

[0183] As will be apparent to those skilled in the art, many different ways are possible to perform the methods described herein (e.g., methods 400 and / or 500). For example, the steps may be performed in the order shown, but the order of the steps may also vary, or some steps may be performed in parallel. Furthermore, other method steps may be inserted between the steps. The inserted steps may represent a refinement of the method described herein, or may be unrelated to the method. Some parts may be performed at least partially in parallel. Moreover, a given part may not be fully completed before the next step begins.

[0184] Embodiments of this method can be executed using software that includes instructions for causing a processor system to perform methods 400 and / or 500. The software may include only those steps taken by a specific sub-entity of the system. The software may be stored on a suitable storage medium (such as a hard disk, floppy disk, memory, optical disk, etc.). The software may be transmitted as a signal along a wire or wirelessly, or using a data network (e.g., the Internet). The software may be made available for download and / or for remote use on a server. Embodiments of this method can be executed using a bitstream arranged to configure programmable logic (e.g., a field-programmable gate array (FPGA)).

[0185] It should be understood that the currently disclosed subject matter also extends to computer programs applicable to putting the currently disclosed subject matter into practice, particularly computer programs on or in a carrier. Programs can be in the form of source code, object code, intermediate source code, and object code such as partially compiled form, or any other form suitable for implementing embodiments of the method. Embodiments relating to computer program products include computer-executable instructions corresponding to each processing step in the processing steps of at least one of the illustrated methods. These instructions may be subdivided into subroutines and / or stored in one or more files that can be statically or dynamically linked. Another embodiment relating to computer program products includes computer-executable instructions corresponding to each device, unit, and / or portion of at least one system and / or product illustrated.

[0186] Figure 7a A computer-readable medium 1000 is illustrated, having a writable portion 1010 including a computer program 1020. The computer program 1020 includes instructions for causing a processor system to execute a segmentation method and / or an atlas selection method according to embodiments. The computer program 1020 may be embodied on the computer-readable medium 1000 as a physical mark or by magnetization of the computer-readable medium 1000. However, any other suitable embodiments are contemplated. Furthermore, it should be understood that although the computer-readable medium 1000 is shown herein as an optical disc, the computer-readable medium 1000 may be any suitable computer-readable medium, such as a hard disk, solid-state storage, flash memory, etc., and may be non-recordable or recordable. The computer program 1020 includes instructions for causing a processor system to execute the segmentation method and / or the atlas selection method.

[0187] Figure 7b A schematic diagram of a processor system 1140 is shown, illustrating an embodiment of selecting devices and / or systems based on segmented devices and / or systems and / or atlases. The processor system includes one or more integrated circuits 1110. Figure 7bThe diagram schematically illustrates the architecture of one or more integrated circuits 1110. Circuit 1110 includes a processing unit 1120, such as a CPU, for running computer program components to perform methods according to embodiments and / or implement modules or units thereof. Circuit 1110 includes a memory 1122 for storing program code, data, etc. A portion of the memory 1122 may be read-only. Circuit 1110 may include a communication element 1126, such as an antenna, a connector, or both, etc. Circuit 1110 may include a dedicated integrated circuit 1124 for performing some or all of the processes defined in the method. Processor 1120, memory 1122, dedicated IC 1124, and communication element 1126 may be interconnected via interconnection components 1130 (such as a bus). Processor system 1110 may be arranged for contact communication and / or contactless communication using antennas and / or connectors, respectively.

[0188] For example, in one embodiment, processor system 1140 (e.g., a partitioning device and / or system and / or atlas selection device and / or system) may include processor circuitry and memory circuitry, the processor being arranged to execute software stored in the memory circuitry. For example, the processor circuitry may be an Intel Core i7 processor, an ARM Cortex-R8, etc. In one embodiment, the processor circuitry may be an ARM Cortex M0. The memory circuitry may be ROM circuitry or non-volatile memory, such as flash memory. Alternatively, the memory circuitry may be volatile memory, such as SRAM memory. In the latter case, the device may include a non-volatile software interface arranged to provide the software, such as a hard disk drive, a network interface, etc.

[0189] Although device 1100 is shown as including one of each described component, in various embodiments, the various components may be repeated. For example, processor 1120 may include multiple microprocessors configured to independently execute the methods described herein, or configured to execute steps or subroutines of the methods described herein, causing multiple processors to cooperate in implementing the functions described herein. Furthermore, in the case where device 1100 is implemented in a cloud computing system, the various hardware components may belong to separate physical systems. For example, processor 1120 may include a first processor in a first server and a second processor in a second server.

[0190] It should be noted that the above embodiments are illustrative and not limiting of the subject matter currently disclosed, and those skilled in the art will be able to devise many alternative embodiments.

[0191] In the claims, any reference marks placed between parentheses should not be construed as limiting the claims. The use of the verb "comprising" and its variations does not exclude the presence of elements or steps other than those described in the claims. The article "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. Expressions such as "at least one of..." when preceding a list of elements indicate the selection of all elements or any subset of elements from the list. For example, the expression "at least one of A, B, and C" should be understood to include only A, only B, only C, both A and B, both A and C, both B and C, or all A, B, and C. The subject matter currently disclosed can be implemented by hardware comprising several completely different elements and by a suitably programmed computer. In device claims that enumerate several parts, several of these parts can be implemented by the same piece of hardware. The fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used advantageously.

[0192] In the claims, reference numerals enclosed in parentheses refer to reference numerals in the accompanying drawings of exemplary embodiments or formulas of embodiments, thereby increasing the comprehensibility of the claims. These reference numerals should not be construed as limiting the claims.

Claims

1. A method (400) for medical image segmentation, comprising: - Obtain (410) a segmentation function (210) configured to receive a medical input image (212) and generate image segmentation (214). - Obtain (520) multiple test images (610) and associated test image segments (620). - The segmentation quality of the plurality of test images is determined (530) by comparing the associated test image segmentation with the image segmentation generated by the segmentation function component. - Select (540) multiple test images from the multiple test images that have a segmentation quality exceeding a threshold as multiple atlas images. - Obtain (420) the multiple atlas images (320). - Receive (430) medical images (312), - Register (440) the received image (312) to the plurality of atlas images to obtain a plurality of registered images (330) and a plurality of corresponding registration transformations, the plurality of corresponding registration transformations being configured to register the received image on the plurality of atlas images; - The segmentation function is applied (450) to the multiple registered images to obtain multiple registered image segmentation (340). - Apply the inverse of the multiple registration transformations (460) to the image segmentation of the multiple registrations to obtain multiple image segments (350). - Determine (470) the output segment (361) from the plurality of image segments.

2. The method of claim 1, wherein the segmentation function is a machine learning function trained on a plurality of training images and corresponding training image segmentation, and wherein the atlas images have been selected from the plurality of training images.

3. The method according to claim 1, comprising: - For test images that have not yet been selected: - Apply registration from the unselected test image (610') to the selected test image (651), and determine the registration quality by comparing the registered test image with the selected test image (640). - Select a test image with registration quality below the threshold as another atlas image.

4. The method according to claim 3, comprising: - For test images that have not yet been selected: - Apply registration from the test image to the selected test image and determine multiple registration qualities; - Determine the overall registration quality based on the determined multiple registration qualities; - Select a test image with overall registration quality below the threshold as another atlas image.

5. The method according to any one of claims 3-4, wherein the registration is any one of the following: translation registration, rigid registration, similarity registration, affine registration, or non-rigid registration.

6. The method according to any one of claims 1-4, wherein determining the segmentation quality and / or registration quality includes determining the Dice score.

7. The method according to any one of claims 1-4, wherein the segmentation function has been trained on a plurality of training images, and the test image is included in the training images.

8. The method according to any one of claims 1-4, wherein the segmentation functional component is a neural network.

9. A medical image segmentation system, comprising: - A segmentation function component storage device, configured to store segmentation function components, said segmentation function components being configured to receive medical input images and generate image segmentation. - An atlas storage device, configured to store multiple atlas images. - The communication interface is configured to receive medical images. - Processor system, configured for: - Obtain multiple test images and associated test image segments. - The segmentation quality of the plurality of test images is determined by comparing the associated test image segments with image segments generated by the segmentation function unit. - Select multiple test images from the multiple test images that have a segmentation quality exceeding a threshold as the multiple atlas images. - Register the received image to the plurality of atlas images to obtain a plurality of registered images and a plurality of corresponding registration transformations, wherein the plurality of corresponding registration transformations are configured to register the received image on the plurality of atlas images; - The segmentation function is applied to the multiple registered images to obtain image segmentation of the multiple registered images; - Apply the inverse of the multiple registration transformations to the image segmentation of the multiple registrations to obtain multiple image segments; - Determine the output segment from the plurality of image segments.

10. The system according to claim 9, further comprising: - A test image storage device configured to store the plurality of test images and the associated test image segments.

11. A transient or non-transient computer-readable medium (1000) comprising data (1020) representing instructions that, when executed by a processor system, cause the processor system to perform the method according to any one of claims 1-8.

Images