Image processing apparatus, method and program, and learning apparatus, method and program

Abstract

In the present invention, a processor derives a reduced image by reducing a target image, extracts from the reduced image the region of a target structure so as to derive a reduced structure image including the region of the target structure, extracts from the target image a corresponding image corresponding to the reduced structure image, and inputs the corresponding image and the reduced structure image into an extraction model created by machine learning of a neural network, so as to extract from the extraction model the region of the target structure included in the corresponding image.

Description

Technical Field

[0001] The present disclosure relates to an image processing apparatus, method, and program, as well as a learning apparatus, method, and program.

Background Art

[0002] In recent years, machine learning techniques using deep learning have attracted attention. In particular, various models for segmenting an object included in an image have been proposed by learning a convolutional neural network (hereinafter referred to as CNN), which is one of multi-layer neural networks in which a plurality of processing layers are hierarchically connected, by deep learning. Further, a method for classifying the segmented regions has also been proposed. For example, Japanese Unexamined Patent Application Publication No. 2019-021313 proposes a method of normalizing an input image, extracting a given region from the normalized image, and applying the extracted region to the input image to classify an object in the given region in the input image.

Summary of the Invention

Problems to be Solved by the Invention

[0003] However, the method described in Japanese Unexamined Patent Application Publication No. 2019-021313 cannot extract an object included in an input image quickly and accurately.

[0004] The present disclosure has been made in view of the above circumstances, and an object thereof is to enable an object included in an image to be extracted quickly and accurately.

Means for Solving the Problems

[0005] The image processing apparatus according to the present disclosure includes at least one processor, The processor derives a reduced image by reducing a target image, The processor derives a reduced structure image including a region of a target structure by extracting a region of the target structure from the reduced image, Extract the corresponding image from the target image that corresponds to the reduced structure image. By inputting the corresponding image and the reduced-size structure image into an extraction model constructed using machine learning on a neural network, the region of the target structure contained in the corresponding image is extracted from the extraction model.

[0006] Furthermore, in the image processing apparatus according to this disclosure, the extraction model consists of multiple processing layers that perform convolution processing, and the input layer has two channels. The processor derives an enlarged image of the structure by enlarging the reduced structure image to the same size as the corresponding image. Alternatively, the two channels in the input layer of the extraction model may be input to the enlarged image of the structure and the corresponding image, respectively.

[0007] Furthermore, in the image processing apparatus according to this disclosure, the neural network consists of multiple processing layers that perform convolutional processing, and the processing layer that processes an image with the same resolution as the reduced structure image has an additional channel for inputting the reduced structure image. The processor may also input the reduced structure image into an additional channel.

[0008] Furthermore, in the image processing apparatus according to this disclosure, the processor divides the region of the target structure extracted from the reduced image, derives a divided reduced structure image that includes each of the divided regions of the target structure, From the corresponding image, multiple corresponding segmented images are derived that correspond to each segmented and reduced structural image. Alternatively, the region of the target structure included in the corresponding image may be extracted on a per-divided corresponding image and per-divided reduced structure image basis.

[0009] The learning device described herein comprises at least one processor, The processor uses a first image containing the region of the target structure extracted from a reduced image of the original image containing the target structure, a second image corresponding to the first image extracted from the original image, and ground truth data representing the extraction result of the target structure from the second image as training data to machine-learn a neural network. As a result, when a reduced image of the target structure derived from a reduced image of the target image containing the target structure and a corresponding image corresponding to the reduced image of the target structure extracted from the target image are input, the processor constructs an extraction model that extracts the region of the target structure from the corresponding image.

[0010] The image processing method disclosed herein derives a reduced image by reducing the size of the target image, By extracting the region of the target structure from the reduced image, a reduced image of the structure containing the region of the target structure is derived. Extract the corresponding image from the target image that corresponds to the reduced structure image. By inputting the corresponding image and the reduced-size structure image into an extraction model constructed using machine learning on a neural network, the region of the target structure contained in the corresponding image is extracted from the extraction model.

[0011] The learning method disclosed herein involves machine learning a neural network using training data, with a first image containing the region of the target structure extracted from a reduced image of the original image containing the target structure, a second image corresponding to the first image extracted from the original image, and ground truth data representing the extraction result of the target structure from the second image. This process constructs an extraction model that extracts the region of the target structure from the corresponding image when a reduced image of the target structure derived from a reduced image of the target image containing the target structure and a corresponding image that corresponds to the reduced image of the target structure extracted from the target image are input.

[0012] Furthermore, the image processing method and learning method described herein may be provided as a program for causing a computer to execute them. [Effects of the Invention]

[0013] According to this disclosure, objects contained in an image can be extracted quickly and accurately.

Brief Description of the Drawings

[0014] [Figure 1] Figure showing the schematic configuration of a diagnostic support system to which an image processing apparatus and a learning apparatus according to an embodiment of the present disclosure are applied [Figure 2] Figure showing the schematic configuration of an image processing apparatus and a learning apparatus according to the present embodiment [Figure 3] Functional configuration diagram of an image processing apparatus and a learning apparatus according to the present embodiment [Figure 4] Figure schematically showing the processing performed in the present embodiment [Figure 5] Figure schematically showing the configuration of an extraction model [Figure 6] Figure showing teacher data used for learning [Figure 7] Figure schematically showing another configuration of an extraction model [Figure 8] Figure showing the display screen of a target image [Figure 9] Flowchart showing the learning process performed in the present embodiment [Figure 10] Flowchart showing the image processing performed in the present embodiment [Figure 11] Figure for explaining the receptive field for the liver region [Figure 12] Figure for explaining the receptive field for the liver region [Figure 13] Figure for explaining the division of the liver region

Modes for Carrying Out the Invention

[0015] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. First, the configuration of a medical information system to which an image processing apparatus and a learning apparatus according to the present embodiment are applied will be described. FIG. 1 is a diagram showing the schematic configuration of the medical information system. The medical information system shown in FIG. 1 includes a computer 一 that incorporates an image processing apparatus and a learning apparatus according to the present embodiment, a photographing apparatus 2, and an image storage server 3, which are connected in a communicable state via a network 4.

[0016] Computer 1 incorporates the image processing device and learning device according to this embodiment, and has the image processing program and learning program of this embodiment installed on it. Computer 1 may be a workstation or personal computer directly operated by the physician performing the diagnosis, or it may be a server computer connected to them via a network. The image processing program and learning program are stored in a storage device of the server computer connected to the network, or in network storage, in a state that is accessible from the outside, and are downloaded and installed on Computer 1 used by the physician as needed. Alternatively, they may be recorded on a recording medium such as a DVD (Digital Versatile Disc) or CD-ROM (Compact Disc Read Only Memory) and distributed, and then installed on Computer 1 from that recording medium.

[0017] The imaging device 2 is a device that generates a three-dimensional image representing a part of the subject to be diagnosed by imaging that part of the subject. Specifically, it is a CT (Computed Tomography) device, an MRI (Magnetic Resonance Imaging) device, or a PET (Positron Emission Tomography) device. The three-dimensional image, consisting of multiple slice images, generated by this imaging device 2 is transmitted to the image storage server 3 and stored. In this embodiment, the imaging device 2 is a CT device, and it generates a three-dimensional image from a CT image of the subject's chest and abdomen.

[0018] The image storage server 3 is a computer that stores and manages various types of data, and is equipped with a large-capacity external storage device and database management software. The image storage server 3 communicates with other devices via a wired or wireless network 4 to send and receive image data, etc. Specifically, it acquires various types of data, including image data of 3D images generated by the imaging device 2, via the network, and stores and manages them on a recording medium such as a large-capacity external storage device. The storage format of the image data and communication between each device via the network 4 are based on protocols such as DICOM (Digital Imaging and Communication in Medicine). The image storage server 3 also stores training data, which will be described later.

[0019] Next, the image processing apparatus and learning apparatus according to this embodiment will be described. Figure 2 illustrates the hardware configuration of the image processing apparatus and learning apparatus according to this embodiment. As shown in Figure 2, the image processing apparatus and learning apparatus (hereinafter sometimes referred to simply as the image processing apparatus) 20 includes a CPU (Central Processing Unit) 11, non-volatile storage 13, and memory 16 as a temporary storage area. The image processing apparatus 20 also includes a display 14 such as a liquid crystal display, input devices 15 such as a keyboard and mouse, and a network interface 17 connected to a network 4. The CPU 11, storage 13, display 14, input devices 15, memory 16, and network interface 17 are connected to a bus 18. The CPU 11 is an example of a processor in this disclosure.

[0020] Storage 13 is implemented using HDD (Hard Disk Drive), SSD (Solid State Drive), flash memory, etc. The image processing program 12A and the learning program 12B are stored in storage 13 as a storage medium. The CPU 11 reads the image processing program 12A and the learning program 12B from storage 13, expands them into memory 16, and executes the expanded image processing program 12A and the learning program 12B.

[0021] Next, the functional configuration of the image processing device and learning device according to this embodiment will be described. Figure 3 is a diagram showing the functional configuration of the image processing device and learning device according to this embodiment. As shown in Figure 3, the image processing device 20 includes an information acquisition unit 21, a reduction unit 22, a first extraction unit 23, a second extraction unit 24, a third extraction unit 25, a learning unit 26, and a display control unit 27. When the CPU 11 executes the image processing program 12A, the CPU 11 functions as the information acquisition unit 21, the reduction unit 22, the first extraction unit 23, the second extraction unit 24, the third extraction unit 25, and the display control unit 27. When the CPU 11 executes the learning program 12B, the CPU 11 functions as the learning unit 26. In this embodiment, the target image G0 is a CT image including the chest and abdomen of a human body, and the liver region is extracted from the target image G0 as the target structure.

[0022] The information acquisition unit 21 acquires the target image G0 to be processed from the image storage server 3 based on instructions from the operator via the input device 15. The information acquisition unit 21 also acquires training data from the image storage server 3 for training the extraction model, which will be described later.

[0023] The processes performed by the reduction unit 22, the first extraction unit 23, the second extraction unit 24, and the third extraction unit 25 will be explained below with reference to Figure 4.

[0024] The reduction unit 22 reduces the target image G0 to derive the reduced image GS0. The reduction ratio can be, for example, 1 / 4, but is not limited to this. For example, it can be between 1 / 2 and 1 / 16.

[0025] The first extraction unit 23 extracts the liver region as the region of the target structure from the reduced image GS0. In this embodiment, the first extraction unit 23 extracts the liver region from the reduced image GS0 using an extraction model 23A constructed by machine learning a neural network. The extraction model 23A consists of a neural network that has been machine-learned to extract the liver region from a CT image when a CT image including the chest and abdomen of the human body is input. In Figure 4, hatching is applied to the extracted liver region in the reduced image GS0.

[0026] In addition to the extraction model 23A, other methods can be used to extract the liver region from the reduced image GS0, including threshold processing based on the voxel values ​​of the target image G0, region growing based on seed points representing the liver region, template matching based on the shape of the liver, and graph cuts.

[0027] The first extraction unit 23 then derives a reduced liver image GS1 by clipping a rectangular region containing the liver region in the reduced image GS0.

[0028] The second extraction unit 24 extracts the region corresponding to the reduced liver image GS1 from the target image G0 as the corresponding image. Specifically, the second extraction unit 24 enlarges the reduced liver image GS1 to the same resolution as the target image G0, and extracts the region in the target image G0 that has the greatest correlation with the enlarged reduced liver image GS1 as the corresponding image G1.

[0029] The third extraction unit 25 extracts the liver region contained in the corresponding image G1. To this end, the third extraction unit 25 has an extraction model 25A constructed by machine learning a neural network to extract the liver region contained in the corresponding image G1 when the corresponding image G1 and the reduced liver image GS1 are input.

[0030] Figure 5 is a schematic diagram showing the configuration of the extraction model 25A. As shown in Figure 5, in this embodiment, the extraction model 25A of the third extraction unit 25 is composed of a U-Net, which is a type of convolutional neural network (CNN). The U-Net shown in Figure 5 consists of seven layers: an input layer 30, the first to fifth layers 31 to 35, and an output layer 36.

[0031] In this embodiment, the third extraction unit 25 derives a reduced liver image GS2 by enlarging the reduced liver image GS1 to the same resolution as the corresponding image G1. The third extraction unit 25 then inputs the corresponding image G1 and the enlarged reduced liver image GS2 to the extraction model 25A. For this purpose, the input layer 30 has a channel 30A into which the corresponding image G1 is input and a channel 30B into which the reduced liver image GS2 is input. Note that the enlarged reduced liver image GS2 may be the enlarged reduced liver image derived by the second extraction unit 24 when extracting the corresponding image G1.

[0032] The input layer 30 concatenates, or convolves, the corresponding image G1 and the enlarged reduced liver image GS2 using a predetermined kernel, and outputs a feature map F1 that integrates the corresponding image G1 and the enlarged reduced liver image GS2. The feature map F1 is input to the first layer 31. In this embodiment, for example, a 3x3 kernel is used for convolution, but it is not limited to this.

[0033] The first layer 31 has, for example, four convolutional layers. The feature map F2 output from the first layer 31 is input to the fifth layer 35. The feature map F2 is then pooled and reduced in size to half before being input to the second layer 32. In the pooling process, the maximum value of the four pixels is used, but this is not the only option.

[0034] The second layer 32 has, for example, four convolutional layers. The feature map F3 output from the second layer 32 is input to the fourth layer 34. The feature map F3 is also pooled and reduced in size to half before being input to the third layer 33.

[0035] The third layer 33 has, for example, eight convolutional layers. The feature map F4 output from the third layer 33 is upsampled to double its size and input to the fourth layer 34.

[0036] The fourth layer 34, for example, has four convolutional layers and performs a convolution operation by integrating the feature map F3 from the second layer 32 and the upsampled feature map F4 from the third layer 33. The feature map F5 output from the fourth layer 34 is upsampled to double its size and input to the fifth layer 35.

[0037] The fifth layer 35, for example, has two convolutional layers and performs a convolution operation by integrating the feature map F2 from the first layer 31 and the upsampled feature map F5 from the fourth layer 34. The feature map F6 output from the fifth layer 35 is input to the output layer 36.

[0038] Output layer 36 outputs extracted image G2, in which the liver region is extracted from the corresponding image G1.

[0039] The extraction model 25A is constructed by machine learning a neural network using a large amount of training data. The training unit 26 performs the training of the neural network. Figure 6 shows an example of the training data used for training. As shown in Figure 6, the training data 40 consists of a training reduced liver image 41 generated by extracting the liver from a reduced image obtained by reducing the medical image, a training corresponding image 42 obtained by extracting the region corresponding to the training reduced liver image 41 in the medical image, and ground truth data 43. The training reduced liver image 41 is an image in which the liver region is masked in the reduced image. The ground truth data 43 is an image in which the liver region in the training corresponding image 42 is masked. The training reduced liver image 41 and the ground truth data 43 are generated by the user extracting the liver region while viewing the image. Note that the training reduced liver image 41 is an example of the first image and the training corresponding image 42 is an example of the second image.

[0040] The learning unit 26 inputs the reduced-size liver image 41 and the corresponding image 42 into the neural network and causes the neural network to extract the liver region from the corresponding image 42. The learning unit 26 then derives the difference between the extraction result by the neural network and the ground truth data 43 as a loss, and learns the connection weights and kernel coefficients of the neural network so that the loss is below a predetermined threshold.

[0041] The learning unit 26 then repeatedly performs learning until the loss falls below a predetermined threshold. As a result, when the reduced liver image GS1 and the corresponding image G1 are input, an extraction model 25A is constructed that extracts the liver region contained in the corresponding image G1. The learning unit 26 may also repeat learning a predetermined number of times.

[0042] Note that the configuration of the U-Net constituting the extraction model 25A is not limited to that shown in Figure 6. For example, as shown in Figure 7, the input layer 30 may be configured with only one channel 30A for inputting the corresponding image G1, and a channel 33A for inputting the reduced liver image GS1 may be added to the third layer 33, which processes the feature map F3 of the same size as the reduced liver image GS1. In this case, the third layer 33 will combine the feature map F3 and the reduced liver image GS1 input to channel 33A and convolve them, outputting a feature map F4 that integrates the feature map F3 and the reduced liver image GS1.

[0043] The display control unit 27 displays the target image G0, from which the liver region has been extracted, on the display 14. Figure 8 shows the display screen of the target image. As shown in Figure 8, the target image G0 is displayed on the display screen 50. In addition, a mask 60 based on the extracted image G2 is applied to the liver region of the displayed target image G0.

[0044] Next, the processes performed in this embodiment will be described. Figure 9 is a flowchart of the learning process performed in this embodiment. It is assumed that the training data is obtained from the image storage server 3 and stored in the storage 13. The learning unit 26 obtains the training data stored in the storage 13 (step ST1) and performs training on the U-Net using the training data (step ST2). This constructs the extraction model 25A.

[0045] Figure 10 is a flowchart showing the image processing performed in this embodiment. It is assumed that the target image G0 is acquired from the image storage server 3 and stored in the storage 13. First, the reduction unit 22 reduces the target image G0 to derive a reduced image GS0 (step ST11). Next, the first extraction unit 23 extracts the liver region from the reduced image GS0 (step ST12) and derives a reduced liver image GS1 by clipping a rectangular region containing the liver region in the reduced image GS0 (step ST13).

[0046] Next, the second extraction unit 24 extracts the region corresponding to the reduced liver image GS1 from the target image G0 as the corresponding image G1 (step ST14). Then, the third extraction unit 25 extracts the liver region from the corresponding image G1 (step ST15). Furthermore, the display control unit 27 displays the target image G0 from which the liver region has been extracted on the display 14 (step ST16), and the process ends.

[0047] Here, we consider extracting the liver region from the corresponding image G1, given no information about the liver region. In this case, as shown in Figure 11, it is necessary to deepen the hierarchy of the neural network constituting the extraction model 25A (i.e., increase the number of layers) in order to obtain a large receptive field 51 that can extract features not only from the boundary of the liver region but also from the inside to the outside of the liver region.

[0048] However, increasing the depth of the neural network increases the processing time for training and extraction, and also requires more memory for processing. Furthermore, it necessitates more training data.

[0049] Furthermore, as shown in the corresponding image G1, when attempting to extract the liver region from a portion of the human body, a significant amount of information surrounding the liver region is lost. This makes it difficult for conventional neural networks to learn, potentially resulting in inaccurate extraction of the liver region.

[0050] In this embodiment, the corresponding image G1 and the reduced liver image GS1 are input to the extraction model 25A to extract the liver region contained in the corresponding image G1. Here, the reduced liver image GS1 provides a rough extraction result of the liver region contained in the corresponding image G1. Therefore, it is sufficient to train the extraction model 25A to distinguish only the boundary portion between the liver and other regions contained in the corresponding image G1. That is, as shown in Figure 12, it is sufficient to obtain a small receptive field 52 that is small enough to extract features around the boundary between the liver region and other regions. As a result, it is possible to reduce the number of layers in the neural network that constitutes the extraction model 25A, and a large amount of memory for processing is not required. Consequently, it becomes possible to extract the liver region from the corresponding image G1 quickly and accurately.

[0051] In the above embodiment, the first extraction unit 23 may divide the extracted liver region and derive a divided reduced liver image containing each of the divided liver regions. Figure 13 is a diagram illustrating the division of the liver region. As shown in Figure 13, the first extraction unit 23 divides the liver region extracted from the reduced image GS0 into an upper region and a lower region to derive a first reduced liver image GS11 and a second reduced liver image GS12. In this case, the second extraction unit 24 extracts a first corresponding image G11 corresponding to the first reduced liver image GS11 and a second corresponding image G12 corresponding to the second reduced liver image GS12 from the target image G0. The first reduced liver image GS11 and the second reduced liver image GS12 are examples of divided reduced structure images, and the first corresponding image G11 and the second corresponding image G12 are examples of divided corresponding images.

[0052] Furthermore, the third extraction unit 25 inputs the first reduced liver image GS11 and the first corresponding image G11 into the extraction model 25A and extracts the upper liver region from the first corresponding image G11. The third extraction unit 25 also inputs the second reduced liver image GS12 and the second corresponding image G12 into the extraction model 25A and extracts the lower liver region from the second corresponding image G12.

[0053] By dividing the liver region into upper and lower regions in this way, especially for the lower liver region, it becomes unnecessary to process the right-side region of the liver compared to using the corresponding image G1 and reduced liver image GS1. Therefore, the computational load on the extraction model 25A can be reduced, and as a result, the extraction of the liver region can be performed more quickly.

[0054] In this case, when dividing the liver region, it is preferable to train the extraction model 25A using training data in which the division method is varied in various ways. This improves the robustness of the extraction model 25A when extracting the liver region from the corresponding image G1 when the liver region is divided.

[0055] In the above embodiment, the liver is used as the object included in the target image G0, but the object is not limited to the liver. In addition to the liver, any part of the human body such as the heart, lungs, brain, and limbs can be used as the object.

[0056] Furthermore, although a CT image is used as the target image G0 in the above embodiment, it is not limited to this. Any image can be used as the target image G0, including 3D images such as MRI images, or radiographic images obtained by simple radiography.

[0057] Furthermore, in the above embodiment, the hardware structure of the Processing Unit that executes various processes such as the information acquisition unit 21, the reduction unit 22, the first extraction unit 23, the second extraction unit 24, the third extraction unit 25, the learning unit 26, and the display control unit 27 can be any of the following types of processors. As mentioned above, these types of processors include a CPU, which is a general-purpose processor that executes software (programs) and functions as various processing units, as well as a Programmable Logic Device (PLD), which is a processor whose circuit configuration can be changed after manufacturing, such as an FPGA (Field Programmable Gate Array), and a dedicated electrical circuit, which is a processor with a circuit configuration specifically designed to execute a particular process, such as an ASIC (Application Specific Integrated Circuit).

[0058] A single processing unit may consist of one of these various processors, or it may consist of a combination of two or more processors of the same or different types (for example, a combination of multiple FPGAs or a combination of a CPU and an FPGA). Alternatively, multiple processing units may be composed of a single processor.

[0059] Examples of configuring multiple processing units with a single processor include, firstly, a configuration where one or more CPUs and software combine to form a single processor, as exemplified by client and server computers, and this processor functions as multiple processing units. Secondly, a configuration using a processor that realizes the functions of the entire system, including multiple processing units, on a single IC (Integrated Circuit) chip, as exemplified by System-on-a-Chip (SoC). Thus, various processing units are configured, in terms of hardware structure, using one or more of the above-mentioned processors.

[0060] Furthermore, the hardware structure of these various processors can more specifically utilize electrical circuits (Circuitry) that combine circuit elements such as semiconductor elements. [Explanation of symbols]

[0061] 1 Computer 2 Imaging device 3. Image storage server 4 Network 11 CPU 12 memory 13 Storage 14 displays 15 Input Devices 20 Image Processing Devices 21 Information Acquisition Department 22 Reduction section 23 1st extraction part 23A Extraction Model 24 Second extraction part 25 Third extraction part 25A Extraction Model 26 Learning Department 27 Display Control Unit 30 Input Layers 30A, 30B channels 31 1st layer 32 2nd layer 33 3rd layer 33A Channel 34 4th layer 35 5th layer 36 Output Layers 40 Training data 41 Teacher-reduced liver images 42 Teacher-Corresponding Images 43 Correct Data 50 display screen 51, 52 Receptive field 60 masks G0 Target Image G1, G11, G12 compatible images G2 Extracted Image GS0 reduced image GS1, GS11, GS12 reduced liver images GS2 Enlarged and reduced liver image

Claims

1. Equipped with at least one processor, The aforementioned processor, By reducing the size of the target image, a reduced image is derived. The region of the target structure is extracted from the reduced image, and the region of the target structure extracted from the reduced image is divided to derive a divided reduced structure image that includes each of the divided regions of the target structure. Multiple corresponding divided images are extracted from the target image, each corresponding to one of the divided and reduced structure images. An image processing device that inputs the segmented image and the segmented reduced structure image into an extraction model constructed by machine learning a neural network, thereby extracting the region of the target structure included in the segmented image from the extraction model on a segmented image and segmented reduced structure image basis.

2. The extraction model consists of multiple processing layers that perform convolution processing, and the input layer has two channels. The processor derives an enlarged image of the divided structure by enlarging the divided and reduced structure image to the same size as the corresponding divided image. The image processing apparatus according to claim 1, wherein the enlarged image of the divided structure and the corresponding divided image are input to the two channels of the input layer of the extraction model, respectively.

3. The neural network consists of multiple processing layers that perform convolutional processing, and the processing layer that processes images with the same resolution as the divided and reduced structure image has an additional channel for inputting the divided and reduced structure image. The image processing apparatus according to claim 1, wherein the processor inputs the divided and reduced structure image to the additional channel.

4. Equipped with at least one processor, The aforementioned processor, A learning device that constructs an extraction model for extracting the divided regions of a target structure from a corresponding divided image when a divided reduced-size image of a target image containing the target structure is input, by machine learning a neural network using a first image containing each of the divided regions of the target structure, a second image corresponding to the first image extracted from the original image, and ground truth data representing the extraction results of the divided target structure from the second image as training data.

5. By reducing the size of the target image, a reduced image is derived. The region of the target structure is extracted from the reduced image, and the region of the target structure extracted from the reduced image is divided to derive a divided reduced structure image that includes each of the divided regions of the target structure. Multiple corresponding divided images are extracted from the target image, each corresponding to one of the divided and reduced structure images. An image processing method that involves inputting the segmented image and the segmented reduced structure image into an extraction model constructed by machine learning a neural network, thereby extracting the region of the target structure included in the segmented image from the extraction model on a segmented image and segmented reduced structure image basis.

6. A learning method for training a neural network using machine learning, wherein when a divided reduced-size image of a target structure derived by dividing the region of the target structure extracted from a reduced-size image of the target image containing the target structure, and a divided corresponding image extracted from the target image containing the divided reduced-size image of the target structure are input, the method constructs an extraction model that extracts the region of the divided target structure from the divided corresponding image.

7. The procedure for deriving a reduced image by reducing the size of the target image, A procedure for extracting the region of the target structure from the reduced image, dividing the region of the target structure extracted from the reduced image, and thereby deriving a divided reduced structure image that includes each of the divided regions of the target structure, A procedure for extracting multiple corresponding divided images from the aforementioned target image, each corresponding to one of the divided and reduced structure images, An image processing program that causes a computer to perform the following steps: input the segmented image and the segmented reduced structure image into an extraction model constructed by machine learning a neural network, thereby extracting the region of the target structure contained in the segmented image from the extraction model on a segmented image and segmented reduced structure image basis.

8. A learning program that uses a neural network as training data to perform machine learning, thereby causing the computer to execute a procedure to construct an extraction model that extracts the regions of the divided target structure from the corresponding divided images, when input is a divided reduced-size image of the target structure derived by dividing the regions of the target structure extracted from a reduced-size image of the target image containing the target structure, a second image corresponding to the first image extracted from the original image, and ground truth data representing the extraction results of the divided target structure from the second image.

Images