Medical image processing apparatus and medical image processing method

Abstract

To reduce a user's burden related to segmentation.SOLUTION: A medical image processing device includes a segmentation unit, a selection unit, and a display control unit. The segmentation unit executes segmentation for a medical image set including a plurality of medical images, and outputs a first result on the segmentation. The selection unit selects an object image suitable as a correction object of the first result of the plurality of medical images on the basis of a difference between a second result of the segmentation executed in the past and a third result obtained by correcting the second result. The display control unit displays the object image.SELECTED DRAWING: Figure 4

Description

Technical Field

[0001] The embodiments disclosed in this specification and the drawings relate to a medical image processing apparatus and a medical image processing method.

Background Art

[0002] Techniques for extracting biological tissues (e.g., human organs) by performing segmentation processing on medical images are known. However, biological tissues may not be accurately extracted. A user (expert) corrects the region indicating the biological tissue for those on which accurate segmentation has not been performed. For example, when segmentation processing is performed on a plurality of slice images (medical images) constituting a set of medical images acquired by a CT (Computed Tomography) apparatus, the user checks the segmentation results and corrects the region indicating the biological tissue (e.g., human organ) for those on which accurate segmentation has not been performed. The user has to check the results of segmentation performed on each of the plurality of slice images constituting the set of medical images one by one and find out the slice images that need correction from among them. Such work becomes more difficult to appropriately select the slice images that need correction as the number of slice images to be subjected to segmentation processing increases, and the burden on the user increases.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0004] One of the problems to be solved by the embodiments disclosed in this specification and the drawings is to reduce the user's burden related to segmentation. However, the problems to be solved by the embodiments disclosed in this specification and the drawings are not limited to the above problems. The problems corresponding to the effects of each configuration shown in the embodiments described later can also be regarded as other problems.

Means for Solving the Problems

[0005] The medical image processing apparatus according to the embodiment includes a segmentation unit, a selection unit, and a display control unit. The segmentation unit performs segmentation on a medical image set including a plurality of medical images and outputs a first result of the segmentation. The selection unit selects a target image suitable for being a correction target of the first result from among the plurality of medical images based on the difference between the second result of the segmentation executed in the past and the third result obtained by correcting the second result. The display control unit displays the target image.

Brief Description of the Drawings

[0006]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Figure 9

Figure 10

Figure 11

Figure 12

Figure 13

[0007] The medical image processing method according to the present embodiment includes a segmentation step of performing segmentation on a medical image, a selection step of selecting a key image that needs to correct the result of the segmentation based on the difference between the result of the segmentation and the correct data, and a display step of displaying the key image.

[0008] Preferably, in the segmentation step, segmentation is performed on a set of medical images obtained by scanning a plurality of scan positions in an organ region.

[0009] Preferably, it further has an interpolation processing step of correcting the result of the segmentation of the set of medical images by interpolation processing, and the interpolation processing is performed based on the key image selected in the selection step.

[0010] Preferably, the selection step corresponds to the scan positions of the medical image set respectively, calculates a difference value indicating the difference between the segmentation result of the medical images at the plurality of scan positions of the corresponding medical image set and the correct data thereof, and has a difference value calculation step of storing the calculated difference value. Based on the recording of the difference value, a plurality of scan positions are selected from the scan positions of the medical image set as key positions, and the medical images corresponding to the key positions are used as the key images.

[0011] Preferably, the interpolation processing step has a user correction step of obtaining the correct segmentation data of the key image when the segmentation result of the key image is corrected by the user. Using an interpolation algorithm that uses the correct segmentation data of the key image and the key image obtained in the user correction step among the medical image set, correct the segmentation result of the non-key images other than the key image, and obtain the correct segmentation data of the non-key images.

[0012] Preferably, in the selection step, a difference value graph showing the relationship between the difference value and the scan position is drawn based on the recording of the difference value, a plurality of candidate positions are specified for each difference value graph, and the key positions are selected based on all the specified candidate positions.

[0013] Preferably, in the selection step, for each difference value graph, a plurality of scan positions of the medical image set are substituted into the difference value graph, the position with the largest difference value among the scan positions of the medical image set is specified, and the area surrounded by the straight line perpendicular to the horizontal axis, the difference value curve, and the horizontal axis at the position with the largest difference value is a multiple of a preset area value, and all straight lines perpendicular to the horizontal axis are specified. The intersection positions of all these straight lines with the horizontal axis and the position with the largest difference value are set as the candidate positions.

[0014] Preferably, in the interpolation processing step, the result of segmenting the non-key image is interpolated based on the intensity gradient of the boundary of the target tissue in the correct data of the segmentation of the key image.

[0015] Preferably, the segmentation is performed by a neural network, teacher data is generated from each medical image and its correct segmentation data, and the neural network is trained.

[0016] Preferably, in the difference value calculation step, the value obtained by dividing the area of the different part between the segmentation result of the medical image and its correct data by the total area is used as the difference value of the image.

[0017] Preferably, it further includes an organ type determination step of determining the type of the organ corresponding to the medical image set, and in the selection step, the key position is selected based on all the candidate positions of the type.

[0018] The medical image processing apparatus according to the present embodiment includes a segmentation unit that performs segmentation on a medical image, a selection unit that selects a key image that needs to correct the segmentation result based on the difference between the segmentation result and its correct data, and a display control unit that displays the key image.

[0019] (Medical Image and Scan Position) Hereinafter, the medical images used in the present embodiment will be described. The segmentation method of the present embodiment performs segmentation of medical images in units of sets. One set of medical images is, for example, a plurality of scanned images of organs obtained by scanning a plurality of scan positions of a living body organ that are open to each other at a certain interval.

[0020] FIG. 1 is a conceptual diagram showing scanning of a lung region to generate a lung region scan image set. In FIG. 1(a), each horizontal line represents each scan position. By scanning at each scan position, one scan image set showing the cross-sectional structure of the lung region corresponding to each scan position is obtained. FIGS. 1(b), (c), and (d) illustrate scan images corresponding to three scan positions, respectively. Although 14 scan positions are shown in FIG. 1, the number and interval of the scan positions can be set as necessary.

[0021] In the segmentation method of this embodiment, a plurality of sets of medical images are processed. Each medical image set is obtained by scanning a plurality of equidistant scan positions of the same organ region of a patient (which may be the same patient or different patients). The quantity and interval of the scan positions between each medical image set do not have to match.

[0022] (Draft Segmentation) In the semi-automatic segmentation method of medical images, first, draft segmentation of the medical images is performed using AI (Artificial Intelligence) technology or a specialized algorithm. However, generally, the accuracy of the draft segmentation cannot meet the requirements of the medical image segmentation accuracy. Therefore, after the draft segmentation, it is necessary to further correct the draft segmentation result of the medical images.

[0023] The draft segmentation method using a machine learning model is an efficient and highly accurate draft segmentation method. When performing draft segmentation using a machine learning model, the medical image is input into a trained machine learning model including a neural network such as a feed-forward neural network or a convolutional neural network, and information indicating all pixels belonging to the target tissue of the corresponding image is calculated through the operation of the machine learning model. Thereby, the draft segmentation of the image is performed.

[0024] (First Embodiment) The image segmentation method (medical image processing method) according to this embodiment is a segmentation method that improves the semi-automatic segmentation method. In this embodiment, for each image, a "draft segmentation result" and a "corrected segmentation result (correct data)", which is the correction result of the draft segmentation result, are saved. In each of the draft segmentation result and the corrected segmentation result of each image, information indicating all the pixels belonging to the target tissue of the image is recorded. Hereinafter, the segmentation of medical images with the lung as the target of segmentation will be described as an example, but this embodiment is not limited thereto.

[0025] Hereinafter, the first embodiment will be described with reference to the drawings.

[0026] FIG. 2 is a block diagram showing the configuration of the image segmentation apparatus 10. The image segmentation apparatus 10 includes a control unit 110, a display control unit 120, an input processing unit 130, a storage unit 140, a communication unit 150, and a bus (not shown). Note that the image segmentation apparatus 10 is an example of a medical image processing apparatus.

[0027] In the image segmentation apparatus 10, the control unit 110, the display control unit 120, the input processing unit 130, the storage unit 140, and the communication unit 150 are connected to a common bus and exchange information via the bus. Further, the image segmentation apparatus 10 is connected to a display device and an input device, which are external devices, via the communication unit 150.

[0028] The control unit 110 includes a segmentation unit 111, a difference value calculation unit 112, a key position selection unit 113, and an interpolation processing unit 114. Further, the control unit 110 causes the image segmentation apparatus 10 to execute the image segmentation method by executing a program stored in the program storage unit 143 included in the storage unit 140.

[0029] The display control unit 120 controls the images displayed on the display device connected to the image segmentation device 10 in accordance with the instructions from the control unit 110. The display control unit 120 is, for example, a GPU (Graphics Processing Unit). The display device is, for example, a liquid crystal display or a plasma display.

[0030] The input processing unit 130 receives the input from the input device and converts the input information from the input device into computer-readable information. The input processing unit 130 is an interface such as a USB interface, for example, and the input device is, for example, a mouse, a keyboard, or a touch panel.

[0031] The storage unit 140 includes an image storage unit 141 that stores medical images, a segmentation storage unit 142 that stores segmentation information of medical images, a program storage unit 143 that stores programs, a machine learning model storage unit 144 that stores trained machine learning models, and a difference value storage unit 145 that stores difference values associated with the scan positions of each medical image set.

[0032] The communication unit 150 connects the image segmentation device 10 to the external devices, namely the display device and the input device, either wired or wirelessly.

[0033] The segmentation unit 111 performs draft segmentation of the medical images of the corresponding set by inputting each medical image set into the machine learning model, for example, and stores the result of the draft segmentation in the segmentation storage unit 142. The method of draft segmentation is not limited to this.

[0034] The difference value calculation unit 112 calculates a difference value corresponding to the scan position of the medical image of the corresponding set based on the draft segmentation result of each medical image set and the corrected segmentation result which is the correction result of the draft segmentation result, and stores the calculated difference value in the difference value storage unit 145 in association with the scan position. The difference value indicates the difference between the draft segmentation result and the corrected segmentation result of a plurality of images respectively corresponding to each scan position of one medical image set.

[0035] Based on the calculated difference value, the key position selection unit 113 selects a plurality of key positions from the plurality of scan positions of one medical image set. The medical image corresponding to the key position is referred to as a "key image". The key image is used to correct the non-key image in the following interpolation process. The key position selection unit 113 includes a candidate position specifying unit. The candidate position specifying unit generates a difference value graph from the input difference value, and specifies candidate positions which are candidates for a plurality of key positions from each scan position based on the difference value graph. Details of the candidate positions will be described later.

[0036] The interpolation processing unit 114 corrects the draft segmentation result of the remaining images in the corresponding image set by interpolation processing from a part of the images in one image set and their corrected segmentation results, and obtains the corrected segmentation results of the remaining images.

[0037] FIG. 3 is a flowchart showing the process of the image segmentation method according to the present embodiment. Hereinafter, the image segmentation method according to the present embodiment will be described with reference to FIG. 3.

[0038] In step S10, the control unit 110 executes the program stored in the program storage unit 143, starts the execution of the image segmentation method, reads out the machine learning model stored in the machine learning model storage unit 144, and causes the segmentation unit 111 to load the corresponding machine learning model.

[0039] Step S20 includes steps S101 to S112, and the details of steps S101 to S112 will be described later.

[0040] In step S20, segmentation is performed on a set of medical images.

[0041] In step S30, it is determined whether all medical image data has been processed. If it is determined as "Yes", the process is completed; otherwise, the process returns to step S20.

[0042] Hereinafter, the process included in step S20 will be described with reference to FIG. 4.

[0043] In step S101, the control unit 110 reads a set of medical images from the image storage unit 141 and inputs the set of medical images into the segmentation unit 111. The set of medical images is, for example, a data set composed of a plurality of medical images obtained by scanning a lung region.

[0044] In step S102, the segmentation unit 111 performs draft segmentation on the set of medical images and stores the result (the first result) in the segmentation storage unit 142. In other words, the segmentation unit 111 performs segmentation on a set of medical images including a plurality of medical images and outputs the first result of the segmentation.

[0045] Specifically, the segmentation unit 111 inputs the set of medical images into the loaded machine learning model and obtains the segmentation result obtained by the machine learning model. In other words, the segmentation unit 111 performs segmentation on the set of medical images including the plurality of medical images obtained by scanning a plurality of scan positions of an organ region.

[0046] In step S103, the control unit 110 determines whether the current execution is the first execution. If it is determined to be "Yes", the process proceeds to step S104; otherwise, the process proceeds to step S110.

[0047] In step S104, the expert (user) inspects all the images in the medical image set and their draft segmentations, selects a plurality of images with low draft segmentation accuracy and appropriate positional intervals from each other as key images, and records the scan positions (also referred to as slice positions or key positions) corresponding to the plurality of corresponding key images.

[0048] The key images are used to correct non-key images in the following interpolation process. The method of selecting key images is not limited to one. Generally, the lower the accuracy of the draft segmentation of a medical image set, the more key images are required in the following interpolation process. However, more key images require more manual segmentations, so it is desirable to suppress the number of key images as much as possible. Also, regarding the interval between the key positions corresponding to the key images, when the accuracy of the draft segmentation is low, if the distance between two key positions is too long, the draft segmentation results of the non-key images between the two key positions may not be efficiently corrected by the following interpolation process. Also, if the corresponding distance is too short, the user needs to perform segmentations frequently. Therefore, it is desirable to appropriately set the positions of the key positions according to the situation.

[0049] In step S105, the display control unit 120 causes the display device to display the anatomy atlas of the segmentation target (the lung in this embodiment), and marks the key positions at the corresponding positions of the anatomy atlas. FIG. 5 is a conceptual diagram showing the image displayed on the display device in step S105. In FIG. 5, the anatomy atlas of the lung is displayed, and the key positions are marked with horizontal lines in the anatomy atlas. In other words, the display control unit 120 displays the target image.

[0050] In step S106, the draft segmentation result of the key image is corrected by the user. FIG. 6 is a conceptual diagram showing the image displayed on the display device in step S106. As shown in FIG. 6, when the key position marked on the anatomy atlas is clicked, the display control unit 120 highlights the corresponding key position on the display device and displays the current segmentation result of the image (key image) corresponding to the position. Thereafter, the user corrects the segmentation result of the corresponding key image through the input device based on the display of the display device. The input processing unit 130 receives the user input from the input device and updates the segmentation result of the key image according to the user input. In other words, the interpolation processing unit 114 obtains the corrected result of the target image when the first result of the target image is corrected by the user.

[0051] In step S107, the interpolation processing unit 114 corrects the draft segmentation result of the non-key images other than the key image with an interpolation algorithm based on the key image and its corrected segmentation result in the medical image set, obtains the corrected segmentation result of the non-key images, and stores it in the segmentation storage unit 142.

[0052] Referring to FIG. 7, the interpolation process according to this embodiment will be described. FIG. 7 is an explanatory diagram for explaining the interpolation process according to this embodiment. In FIG. 7, the actual states of the target tissues of two key images (the uppermost and the lowermost) and two non-key images sandwiched between the two key images are displayed in the (a) column, and the segmentation states before and after the interpolation process of each image are displayed in the (b) and (c) columns. The parts accurately marked and the unmarked parts in the target tissue are displayed in dark and light colors, respectively. In the (b) column of FIG. 7, since the segmentation results of the key images at the uppermost and the lowermost are corrected by the user, all parts belonging to the target tissue are accurately marked. For the two non-key images in the middle, only some regions belonging to the target tissue are accurately marked in the draft segmentation, so both dark and light parts are provided. In the (c) column of FIG. 7, when the segmentation results of the non-key images in the middle are interpolated based on the segmentation results of the key images at both ends, the two non-key images in the middle will only have dark parts because all parts belonging to the target tissue are accurately marked. In other words, the interpolation processing unit 114 corrects the first result of a medical image different from the target image in the medical image set by interpolation processing based on the correction result of the target image selected in the selection unit.

[0053] The interpolation algorithm used in this embodiment will be described below. The interpolation algorithm of this embodiment searches for the exact boundary of the target tissue in the non-key images based on the intensity gradient of the boundary of the target tissue indicated by the corrected segmentation result of the key images. In medical images, since each tissue exhibits a different intensity, the boundary of each tissue has a specific intensity gradient. In the interpolation algorithm of this embodiment, first, the boundary of the target tissue of each key image is searched from the corrected segmentation result of each key image, and the intensity gradient of the boundary is calculated. Then, for all non-key images sandwiched between each key image and its next key image, the draft segmentation boundary of the target tissue of the non-key image is searched from the draft segmentation result of the non-key image for each non-key image, and within a certain range around the draft segmentation boundary of the target tissue, a boundary whose intensity gradient matches the intensity gradients of the boundaries of the target tissues of the key images at both ends of the non-key image is searched and the draft segmentation boundary is replaced. Then, the draft segmentation result is corrected by determining the pixels included in the target tissue based on the corrected boundary of the target tissue. In other words, the interpolation processing unit 114 interpolates the first result of the medical image different from the target image in the medical image set based on the intensity gradient of the boundary of the target tissue in the corrected result of the target image.

[0054] In addition to the interpolation algorithm based on the intensity gradient of the tissue boundary, the draft segmentation result of the non-key image between the key images may be corrected using an interpolation algorithm related to the shape based on the shape of the target tissue in each key image and its next key image.

[0055] In step S108, the user determines whether the corrected segmentation result of each interpolated medical image meets the desired accuracy. If it is determined as "Yes", the process proceeds to step S109; otherwise, the process returns to step S106. Also, step S108 may not be provided. That is, instead of performing the determination in step S108, the process may proceed to step S109 immediately after performing step S107 without performing the determination in step S108.

[0056] In step S109, the control unit 110 stores the draft segmentation results and the corrected segmentation results of all the images in the segmentation storage unit 142 of the storage unit 140.

[0057] In step S103, when the control unit 110 determines that the execution of the image segmentation method according to the present embodiment this time is not the first execution, that is, when it determines "No", it proceeds to step S110.

[0058] In step S110, the difference value calculation unit 112 calculates the difference value for each scan position based on the draft segmentation result and the corrected segmentation result of the previous medical image set stored in the segmentation storage unit 142, and stores the calculated difference value in the difference value storage unit 145 in association with the scan position. Since the difference value of the previous medical image set at that time is calculated each time step S110 is executed, the difference values of each set so far are stored in the difference value storage unit 145. In the present embodiment, the difference value of each scan position is a value obtained by dividing the area of the different part between the draft segmentation result and the corrected segmentation of the medical image corresponding to the scan position by the total area. The area of the different part is equal to the number of pixels classified into different categories (tissues) between the draft segmentation result and the corrected segmentation result of the corresponding medical image. The total area is the total number of pixels of the target tissue of the corresponding medical image. The method of setting the difference value is not limited to this, and other values representing the difference between the draft segmentation result and the corrected segmentation result may be used. In other words, the difference value calculation unit 112 calculates, as the difference value, a value obtained by dividing the area of the different part between the first result and the corrected result of the first result of each of the plurality of medical images by the total area of the part to be segmented. Note that the difference value calculation unit 112 is an example of a calculation unit.

[0059] In step S111, the candidate position specifying unit of the key position selecting unit 113 draws difference value graphs respectively based on the difference values of all medical image sets stored in the difference value storage unit 145, and specifies a plurality of candidate positions for each difference value graph. FIG. 8 is a conceptual diagram showing the difference value graph of the image segmentation method according to the first embodiment. The horizontal axis represents the continuous positions of the target tissue, i.e., the lungs, in the present embodiment, and the vertical axis represents the difference value. First, the candidate position specifying unit reads out the scan positions of each medical image set so far and the difference values associated with the scan positions from the difference value storage unit 145, and for each medical image set, marks the points corresponding to each scan position and the difference value of the medical image set on the difference value graph, and obtains a difference value curve from those points, for example, by curve fitting. Then, the candidate position specifying unit specifies a plurality of candidate positions for each difference value graph and records those candidate positions. The difference value graph indicates the segmentation accuracy of the draft segmentation at each position within a certain scan interval. In the present embodiment, the difference value graph indicates the segmentation accuracy of the draft segmentation at each position within the lung region. The larger the difference value at a certain position, the lower the draft segmentation accuracy corresponding to the scan position.

[0060] The following describes the method for specifying candidate positions in this embodiment. However, this method is an example, and the method for specifying candidate positions is not limited to this. When the difference value graph is known, the candidate positions are a plurality of positions suitable for the key positions of the corresponding scan section inferred from the difference value graph. FIG. 9 is a diagram for explaining an example of the candidate position specifying method in this embodiment. In the candidate position specifying method of this embodiment, for a certain difference value graph, first, the scan position of the currently processed medical image set is substituted into the difference value graph, the difference values on the difference value curve at each scan position of the medical image set are calculated, and the scan position with the largest difference value is set as one of the candidate positions. In FIG. 9, since the position with the largest difference value is position p1, position p1 is set as the candidate position. Then, at the position where the difference value is the largest, the area surrounded by the vertical line perpendicular to the horizontal axis, the difference value curve, and the horizontal axis is a multiple of a preset area value, and all the vertical lines perpendicular to the horizontal axis are specified, and the intersection positions of all these lines and the horizontal axis are set as candidate positions. Specifically, first, on the left side of position p1, the area surrounded by the vertical line l1 perpendicular to the horizontal axis at position p1, the difference value curve, and the horizontal axis is the preset area value, and the vertical lines perpendicular to the horizontal axis are specified, and the intersection positions of the specified lines and the horizontal axis are set as candidate positions. In FIG. 9, since the area surrounded by the straight line l2, the straight line l1, the difference value curve, and the horizontal axis is the preset area, the position p2 corresponding to the straight line l2 is set as the candidate position. Then, on the left side of the newly specified position p2, the area surrounded by the straight line l2, the difference value curve, and the horizontal axis is the preset area value, and the vertical lines perpendicular to the horizontal axis are specified, and the intersection positions of the lines and the horizontal axis are set as candidate positions. After that, the above process is repeated until no new candidate positions are found. Then, the same process is performed on the right side of the position where the difference value is the largest. When all the candidate positions on the left and right sides are found, those candidate positions are set as the candidate positions of this difference value graph. The above preset area value is set according to the needs of the user. When the user requires high segmentation accuracy, a lower area value is set, and when the user requires low segmentation accuracy, a higher area value is set. Also, the lower the area value, the more images the user needs to manually correct.

[0061] In step S112, the key position selection unit 113 selects a plurality of key positions based on all the candidate positions specified for each difference value graph in step S111. Since each candidate position may be different from the scan position of the medical image set currently being processed, it is necessary to associate each candidate position with each scan position of the medical image set currently being processed. Specifically, the key position selection unit 113 associates all the candidate positions specified in step S111 with the scan position closest to the corresponding candidate position among each scan position of the medical image set currently being processed. After associating all the candidate positions with the scan positions, the probability that each scan position is selected as a candidate position is calculated based on the number of times each scan position is selected as a candidate position and the number of sets of medical images processed so far, and a scan position whose corresponding probability exceeds a preset threshold is selected as a key position. In this embodiment, the threshold is set to, for example, "0.6", that is, in the segmentation of medical images so far, a scan position whose probability of being selected as a candidate position exceeds 60% is selected as a key position.

[0062] In other words, the key position selection unit 113 selects a target image suitable for being a correction target of the first result among a plurality of medical images based on the difference between the second result of the segmentation executed in the past and the third result obtained by correcting the second result. For example, the key position selection unit 113 selects the target image such that the denser the target image is as the difference value indicating the difference between the second result and the third result is larger for each of a plurality of scan positions. Note that the key position selection unit 113 is an example of a selection unit.

[0063] Generally, machine learning models and algorithms used for medical image segmentation output similar segmentation results for similar images (for example, multiple medical images obtained by scanning multiple scan positions within a certain scan range). Therefore, images with low segmentation accuracy often correspond to the same position. In this embodiment, draft segmentation is performed on multiple sets of medical images obtained by scanning multiple scan positions within a certain scan range using the same machine learning model. Therefore, the positions of images suitable for the key images in each set of images (that is, target images suitable for being corrected for the first result), that is, the positions suitable for the key positions, are similar. By statistically analyzing the positions suitable for the key positions in the processing of each set of images up to now, the positions suitable for the key positions of each set of images in the future can be accurately selected.

[0064] According to the image segmentation method of this embodiment, the key position can be automatically and accurately selected to present the key image, so the step of the user manually selecting the key image can be omitted, reducing the burden on the user and improving the efficiency of image segmentation. For example, according to the image segmentation method of this embodiment, medical images that require correction of the segmentation result can be appropriately selected, reducing the burden on the user.

[0065] Also, conventionally, when the user selects a key image, the selection is visually made based on experience and accurate calculation cannot be performed, so there is a risk of selecting an inappropriate medical image as the key image. On the other hand, according to the image segmentation method of this embodiment, the key image is selected based on the difference values of each set of medical images up to now, so the key image can be accurately selected.

[0066] (Second Embodiment) Hereinafter, the second embodiment will be described with reference to the drawings. In the second embodiment, the same or corresponding parts as those in the first embodiment are denoted by the same reference numerals, and redundant descriptions are omitted.

[0067] In this embodiment, a neural network is used as a machine learning model, and a neural network such as a feedforward neural network or a convolutional neural network is used.

[0068] FIG. 10 is a block diagram showing the configuration of an image segmentation apparatus 10B according to the second embodiment. The image segmentation apparatus 10B further includes a training unit 115 in the control unit 110 as compared with the image segmentation apparatus 10.

[0069] The training unit 115 trains a machine learning model based on the result of image segmentation. For example, the training unit 115 uses teacher data to train the neural network stored in the machine learning model storage unit 144. The teacher data consists of a plurality of medical images and their segmentation correct data. The training method used in the training unit 115 is, for example, batch gradient descent method, stochastic gradient descent method, etc.

[0070] FIG. 11 is a flowchart showing the process of the image segmentation method according to the second embodiment. Hereinafter, the image segmentation method according to this embodiment will be described with reference to FIG. 11. The segmentation method of the second embodiment further has step S40B and step S50B as compared with the first embodiment.

[0071] In step S40B, the control unit 110 determines whether a preset number of medical images have been processed. If it is determined as "Yes", the process proceeds to step S50B. Otherwise, the process returns to step S20.

[0072] In step S50B, the training unit 115 generates teacher data from the medical images that have been segmented so far and their segmentation results, trains the neural network stored in the machine learning model storage unit 144, and performs initialization, that is, sets the number of executions of the image segmentation method of this embodiment to 0. Note that the training unit 115 may be provided in the segmentation unit 111.

[0073] In other words, the segmentation unit performs the segmentation by a neural network, generates teacher data based on the first results of each of the plurality of medical images and the corrected results of the first results, and trains the neural network. Here, the corrected results of the first results used here include the corrected results corrected by the user and the corrected results corrected by the interpolation process.

[0074] According to the segmentation method of the present embodiment, after processing a preset number of medical images, the neural network is trained (updated). Therefore, while performing the segmentation of the medical images, the neural network is trained by performing the segmentation of the medical images. As the accuracy of the draft segmentation of the neural network increases, the number of key images selected decreases, so the efficiency of segmentation can be further improved.

[0075] (Third Embodiment) Hereinafter, the third embodiment will be described with reference to the drawings. In the third embodiment, the same reference numerals are given to the same or corresponding parts as those in the first embodiment, and the overlapping description will be omitted.

[0076] In the present embodiment, each medical image set is obtained by scanning a lung region, and the medical images are segmented with the lung as the target of segmentation. Since the lungs of each patient are different, for example, in addition to the lungs of normal adults, there are also lungs whose shapes have changed due to stress and diseased lungs. Therefore, the scanned images of the lungs are classified into a plurality of types.

[0077] FIG. 12 is a block diagram showing the configuration of an image segmentation apparatus 10C according to the third embodiment. The image segmentation apparatus 10C includes a difference value storage unit 145C instead of the difference value storage unit 145 as compared with the image segmentation apparatus 10, and the control unit 110 further includes an organ type determination unit 116.

[0078] The difference value storage unit 145C stores difference values associated with the scan positions of each medical image set for each type.

[0079] The organ type determination unit 116 determines the type of the organ corresponding to a medical image set.

[0080] The difference between the image segmentation method of the third embodiment and the image segmentation method of the first embodiment is that the third embodiment has steps S111C and S112C instead of steps S111 and S112, and further has step S113C. Hereinafter, the description of the steps same as those of the first embodiment is omitted.

[0081] FIG. 13 is a flowchart showing the process of the segmentation steps of the image segmentation method of the third embodiment.

[0082] In step S113C, the organ type determination unit 116 determines the type of the organ (part) corresponding to the medical image set. Note that the type of the part represents, for example, the type of an organ such as "lung", "liver", "brain", etc., but is not limited thereto. For example, the type of the part may represent an organ of a specific subject such as "adult lung", "child lung", "female lung", etc., or may represent a difference in state such as "lung whose shape has changed due to stress", "diseased lung".

[0083] In step S111C, the candidate position specifying unit of the key position selection unit 113 draws a difference value graph for each difference value of each medical image set stored in the difference value storage unit 145 in which the organ type matches the organ type corresponding to the medical image set, and specifies a plurality of candidate positions for each difference value graph.

[0084] In step S112C, the key position selection unit 113 selects a plurality of key positions based on all the candidate positions specified for each difference value graph in which the organ type in step S111 matches the organ type corresponding to the medical image set. In other words, the key position selection unit 113 selects a target image based on the difference between the second result and the third result corresponding to the part of the determined type.

[0085] According to the image segmentation method of the present embodiment, since key positions are selected for each organ type, the selection accuracy of the key positions can be further improved.

[0086] Although several embodiments have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, replacements, changes, and combinations of embodiments can be made without departing from the gist of the invention. These embodiments and their modifications are included in the scope and gist of the invention, and are also included in the invention described in the claims and its equivalent scope.

[0087] According to at least one of the embodiments described above, the burden on the user regarding segmentation can be reduced.

Explanation of Signs

[0088] 10 Image segmentation device 111 Segmentation unit 112 Difference value calculation unit 113 Key position selection unit 114 Interpolation processing unit 115 Training unit 116 Organ type determination unit 120 Display control unit

Claims

1. A segmentation unit that performs segmentation on a medical image set including a plurality of medical images and outputs a first result of the segmentation; A selection unit that selects a target image suitable for being a correction target of the first result from among the plurality of medical images based on a difference between a second result of the segmentation executed in the past and a third result obtained by correcting the second result; A display control unit that displays the target image; A medical image processing apparatus comprising the above.

2. The segmentation unit performs segmentation on the medical image set including the plurality of medical images obtained by scanning a plurality of scan positions of an organ region. The medical image processing apparatus according to Claim 1.

3. The medical image processing apparatus according to Claim 2, further comprising an interpolation processing unit that corrects a first result of a medical image different from the target image in the medical image set by interpolation based on a correction result of the target image selected by the selection unit. The medical image processing apparatus according to Claim 2.

4. For each of the plurality of scan positions, the selection unit selects the target image such that the higher the difference value indicating the difference between the second result and the third result, the denser the target image. The medical image processing apparatus according to Claim 2 or 3.

5. The interpolation processing unit obtains the correction result of the target image when the first result of the target image is corrected by a user. The medical image processing apparatus according to Claim 3.

6. The selection unit draws a plurality of difference value graphs showing the relationship between the difference value and the scan position based on the difference value indicating the difference between the second result and the third result, identifies a plurality of candidate positions for each difference value graph, and selects the target image based on all the identified candidate positions. The medical image processing apparatus according to Claim 5.

7. For each difference value graph, the selection unit substitutes the plurality of scan positions of the medical image set into the difference value graph, identifies the position with the largest difference value among the scan positions of the medical image set, and determines that the area surrounded by a vertical line perpendicular to the horizontal axis, the difference value curve, and the horizontal axis at the position with the largest difference value is a multiple of a preset area value, and identifies all vertical lines perpendicular to the horizontal axis, and sets each intersection position between all those vertical lines and the horizontal axis and the position with the largest difference value as the candidate positions. The medical image processing apparatus according to claim 6.

8. The interpolation processing unit performs interpolation processing on the first result of a medical image different from the target image among the medical image sets based on the intensity gradient of the boundary of the target tissue in the correction result of the target image. The medical image processing apparatus according to any one of claims 5 to 7.

9. The segmentation unit performs the segmentation by a neural network, generates teacher data based on the first result of each of the plurality of medical images and the correction result of the first result, and trains the neural network. The medical image processing apparatus according to any one of claims 5 to 7.

10. The medical image processing apparatus further includes a calculation unit that calculates, as a difference value indicating the difference between the second result and the third result, a value obtained by dividing the area of a different part between the first result of each of the plurality of medical images and the correction result of the first result by the total area of the part to be segmented. The medical image processing apparatus according to any one of claims 5 to 7.

11. The medical image processing apparatus further includes a determination unit that determines the type of the part corresponding to the medical image set, and the selection unit selects the target image based on the difference between the second result and the third result corresponding to the part of the determined type. The medical image processing apparatus according to any one of claims 5 to 7.

12. A medical image processing method including: a segmentation step of performing segmentation on a medical image set including a plurality of medical images and outputting a first result of the segmentation; a selection step of selecting, as a target image suitable for correction of the first result, among the plurality of medical images, based on the difference between the second result of the segmentation executed in the past and the third result obtained by correcting the second result; and a display control step of displaying the target image. ​

Images