MRI image sacroiliac joint segmentation method and device based on convolutional neural network

Abstract

The present application discloses a method for segmenting sacroiliac joints in MRI images based on a convolutional neural network. The method preprocesses an input first MRI image to obtain a second MRI image with higher adaptability, and then segmentes the second MRI image in sequence through a two-stage segmentation method of coarse-grained segmentation and fine-grained segmentation to obtain second target areas in eight different quadrants. The complex image segmentation task is processed step by step, thereby improving the accuracy of the second target area obtained by segmentation. Moreover, the entire process of obtaining the second target area from the first MRI image is completed automatically, and its efficiency is improved compared to obtaining the second target area by manually judging the first MRI image. Moreover, because the entire process is completed automatically, human participation is reduced, so the consistency of the obtained results is improved and will not be interfered by human factors.

Description

Technical Field

[0001] The present application belongs to the field of image information technology, and in particular, relates to a method and device for segmenting sacroiliac joints in MRI images based on a convolutional neural network. Background Art

[0002] In the field of medical image analysis, the segmentation of the sacroiliac joint in Magnetic Resonance Imaging (MRI) images is a very important task because it plays an important role in the differential diagnosis of many diseases. The sacroiliac joint is located between the pelvis and the spine and is responsible for transferring the weight of the upper body to the lower body. Due to its unique anatomical position and the uneven surface of the joint, the sacroiliac joint has become one of the more challenging parts of human body imaging. Accurate segmentation of the sacroiliac joint can provide important anatomical information to help doctors make accurate diagnoses.

[0003] Traditional MRI image sacroiliac joint segmentation methods mainly rely on manual operations, requiring professional doctors to manually draw the joint area. This method is time-consuming and labor-intensive, and is affected by individual differences and the subjective judgment of doctors. The repeatability and accuracy of the results are low. Therefore, an automatic and highly accurate sacroiliac joint segmentation method is needed. Summary of the invention

[0004] The purpose of this application is to provide a convolutional neural network-based sacroiliac joint segmentation method for MRI images, aiming to solve the problems of high reliance on manual labor and low accuracy in sacroiliac joint segmentation in traditional MRI images.

[0005] A first aspect of an embodiment of the present application provides a method for segmenting sacroiliac joints in MRI images based on a convolutional neural network, comprising:

[0006] Performing heterogeneity processing on the input first MRI image to obtain a second MRI image;

[0007] Dividing the second MRI image according to the first label by a first convolutional neural network to obtain a first target area;

[0008] The second MRI image is divided according to the second label and the first target area by a second convolutional neural network to obtain a second target area; wherein the second target area includes eight quadrants of the sacroiliac joint.

[0009] In one optional implementation manner, dividing the second MRI image according to the first label to obtain the first target area includes:

[0010] The second MRI image is divided according to the first label to obtain the target layer number of the sacroiliac joint and the first target area.

[0011] In one optional implementation, the first convolutional neural network is an encoder-decoder architecture; and the training of the first convolutional neural network includes:

[0012] The encoder of the first convolutional neural network processes the input second MRI training image to obtain a first feature map;

[0013] The decoder of the first convolutional neural network processes the first feature map to obtain the first predicted target area;

[0014] Calculate a first error value between the first predicted target area and the first real target area by using DiceLoss;

[0015] L_{ROI1}=DiceLoss(predictive_ROI1, GroundTruth1);

[0016] Where: L_{ROI1} is the first error value; DiceLoss is the loss function; predictive_ROI1 is the first predicted target area; GroundTruth1 is the first true target area;

[0017] The layer number prediction module of the first convolutional neural network processes the first feature map to obtain a first predicted layer number of the sacroiliac joint;

[0018] Calculate a second error value between the first predicted number of layers and the first true number of layers of the sacroiliac joint by using MseLoss;

[0019] L_{Layer1}=(predictive_layer1-true_layer1)^2;

[0020] Where: L_{Layer1} is the second error value; predictive_layer1 is the first prediction layer number;

[0021] true_layer1 is the first true layer number;

[0022] The first error value and the second error value are added to obtain a third error value; and the third error value is gradient-transferred back to adjust the parameters of the first convolutional neural network.

[0023] In one optional implementation, the second convolutional neural network is an encoder-decoder architecture; and the training of the second convolutional neural network includes:

[0024] The encoder of the second convolutional neural network processes the input second MRI training image and the first predicted target area to obtain a second feature map;

[0025] The decoder of the second convolutional neural network processes the second feature map to obtain a second predicted target area;

[0026] Calculate a fourth error value between the second predicted target area and the second true target area by using DiceLoss;

[0027] L_{ROI2}=DiceLoss(predictive_ROI2, GroundTruth2);

[0028] Where: L_{ROI2} is the fourth error value; DiceLoss is the loss function; predictive_ROI2 is the second predicted target area; GroundTruth2 is the second true target area;

[0029] The layer number prediction module of the second convolutional neural network processes the second feature map to obtain a second predicted layer number of the sacroiliac joint;

[0030] Calculate a fifth error value between the second predicted number of layers and the second true number of layers of the sacroiliac joint by using MseLoss;

[0031] L_{Layer2}=(predictive_layer2-true_layer2)^2;

[0032] Where: L_{Layer2} is the fifth error value; predictive_layer2 is the second prediction layer number;

[0033] true_layer2 is the second true layer number;

[0034] The fourth error value and the fifth error value are added to obtain a sixth error value; and the sixth error value is gradient-transferred back to adjust the parameters of the second convolutional neural network.

[0035] In one optional implementation, the layer number prediction module of the first convolutional neural network includes several convolutional layers, pooling layers and linear layers; the layer number prediction module of the second convolutional neural network includes several convolutional layers, pooling layers and linear layers.

[0036] In one optional implementation manner, the performing heterogeneity processing on the input first MRI image includes:

[0037] The size of the first MRI image is adjusted to a preset size, the distribution of voxel intensity values ​​of the first MRI image is truncated within a preset interval, and the truncated voxel intensity values ​​are normalized.

[0038] In one optional implementation, the first convolutional neural network and the second convolutional neural network both adopt a 3D U-Net network.

[0039] A second aspect of the embodiment of the present application provides an MRI image sacroiliac joint segmentation device based on a convolutional neural network, comprising:

[0040] A data preprocessing module, used for performing heterogeneity processing on the input first MRI image to obtain a second MRI image;

[0041] A coarse-grained segmentation module, configured to divide the second MRI image according to the first label by using a first convolutional neural network to obtain a first target area;

[0042] A fine-grained segmentation module is used to divide the second MRI image according to the second label and the first target area through a second convolutional neural network to obtain a second target area; wherein the second target area includes eight quadrants of the sacroiliac joint.

[0043] A third aspect of an embodiment of the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method as described in any one of the above items is implemented.

[0044] A fourth aspect of an embodiment of the present application provides a server, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the methods listed above when executing the computer program.

[0045] Compared with the prior art, the embodiments of the present application have the following beneficial effects: the input first MRI image is preprocessed to obtain a second MRI image with higher adaptability, and then the second MRI image is segmented by a two-stage segmentation method of coarse-grained segmentation and fine-grained segmentation to obtain second target areas in 8 different quadrants, and the complex image segmentation task is processed step by step, thereby improving the accuracy of the second target area obtained by segmentation; and the entire process of obtaining the second target area from the first MRI image is completed automatically, and its efficiency is improved compared to obtaining the second target area by manually judging the first MRI image; and because the entire process is completed automatically, human participation is reduced, so the consistency of the results obtained is improved and will not be interfered by human factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A schematic flowchart of a convolutional neural network-based MRI image sacroiliac joint segmentation method provided in an embodiment of the present application;

[0047] Figure 2 Schematic diagram of the training grid structure of the first convolutional neural network and the second convolutional neural network in an embodiment of the present application. DETAILED DESCRIPTION

[0048] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0049] In recent years, deep learning technology has made significant progress in the field of medical image analysis, especially convolutional neural networks have performed well in image segmentation tasks. Convolutional neural networks have powerful feature extraction and spatial modeling capabilities, and can automatically identify and segment regions of interest in images by learning a large amount of labeled data. Therefore, applying convolutional neural networks to the automatic segmentation of sacroiliac joints is expected to improve the accuracy and efficiency of segmentation. However, there are still some problems with the existing sacroiliac joint segmentation methods based on convolutional neural networks. First, due to the complex anatomical structure of the sacroiliac joint, the existing methods often cannot accurately extract the edge and detail information of the joint; second, the existing segmentation methods often only segment the approximate area of ​​the sacroiliac joint, without segmenting finer quadrant areas.

[0050] Figure 1 The first exemplary principle block diagram provided by the embodiment of the present application is shown. For the convenience of explanation, only the part related to the present embodiment is shown, which is described in detail as follows:

[0051] A convolutional neural network-based sacroiliac joint segmentation method for MRI images, including S100-S300.

[0052] S100: Perform heterogeneity processing on an input first MRI image to obtain a second MRI image.

[0053] Since different MRI image acquisition devices and protocols generally differ, there will be differences between the MRI images output by different acquisition devices. By preprocessing the heterogeneity of the first MRI image, the second MRI image can be more unified, thus more suitable for subsequent image processing, and the accuracy of subsequent image segmentation is improved.

[0054] S200. Divide the second MRI image according to the first label using a first convolutional neural network to obtain a first target area.

[0055] The first convolutional neural network first divides the second MRI image into a first target area through a first label. The first target area is a rough area with a relatively coarse granularity compared to a subsequent second target area.

[0056] S300. Divide the second MRI image according to the second label and the first target area through a second convolutional neural network to obtain a second target area; wherein the second target area includes eight quadrants of the sacroiliac joint.

[0057] The second convolutional neural network then divides the second MRI image according to the first target area and the second label, thereby dividing the second MRI image into second target areas containing 8 different quadrants. When the second convolutional neural network performs segmentation, because the first target area with a reduced range is introduced as a parameter, the difficulty of the second convolutional neural network in processing the second MRI image is reduced, thereby improving the accuracy of obtaining the second target area.

[0058] In this embodiment, the input first MRI image is preprocessed to obtain a second MRI image with higher adaptability, and then the second MRI image is segmented by a two-stage segmentation method of coarse-grained segmentation and fine-grained segmentation to obtain second target areas in 8 different quadrants. The complex image segmentation task is processed step by step, which improves the accuracy of the second target area obtained by segmentation; and the entire process of obtaining the second target area from the first MRI image is completed automatically, and its efficiency is improved compared with obtaining the second target area by manually judging the first MRI image; and because the entire process is completed automatically, human participation is reduced, so the consistency of the obtained results is improved and will not be interfered by human factors.

[0059] In one optional embodiment, the first neural convolutional network and the second neural convolutional network belong to the same neural convolutional network, and both have basically the same structure, but different input channels, wherein the input channel of the first neural convolutional network is only 1, and the input channel of the second neural convolutional network is 2.

[0060] In one optional implementation manner, dividing the second MRI image according to the first label to obtain the first target area includes:

[0061] The second MRI image is divided according to the first label to obtain the target layer number of the sacroiliac joint and the first target area.

[0062] The target layer number of the sacroiliac joint refers to the starting layer number of the sacroiliac joint in the second MRI image.

[0063] The first label is a label of a region mask that meets the preset requirements and is manually annotated, and has 8 quadrants, and the labels of the 8 quadrants are all 1. The second label is the same as the first label in that the label of the region mask that meets the preset requirements and is manually annotated has 8 quadrants; the second label is different from the first label in that the labels of the 8 quadrants are 1, 2, 3, 4, 5, 6, 7, and 8 respectively.

[0064] The regional mask can be understood as a matrix or image of the same size as the first MRI image, which is used to indicate or identify pixels or regions in a specific area of ​​the first MRI image that meets the preset requirements. The regional mask here can be binary, that is, it only contains values ​​of 0 or 1, where a pixel value of 1 indicates that the pixel corresponding to the first MRI image belongs to a specific area or target that meets the preset requirements, and a pixel value of 0 indicates that it does not belong to the area or target. The mask can also be multi-channel, and each channel may represent different types of information or targets.

[0065] Because the first label unifies the labels of the eight quadrants into 1, the first convolutional neural network performs coarse-grained segmentation when segmenting the first target area, and the general area is segmented, which reduces the difficulty.

[0066] See also Figure 2 In one optional implementation, the first convolutional neural network is an encoder-decoder architecture; the training of the first convolutional neural network includes:

[0067] The encoder of the first convolutional neural network processes the input second MRI training image to obtain a first feature map;

[0068] The decoder of the first convolutional neural network processes the first feature map to obtain the first predicted target area;

[0069] Calculate a first error value between the first predicted target area and the first real target area by using DiceLoss;

[0070] L_{ROI1}=DiceLoss(predictive_ROI1, GroundTruth1);

[0071] Where: L_{ROI1} is the first error value; DiceLoss is the loss function; predictive_ROI1 is the first predicted target area; GroundTruth1 is the first true target area;

[0072] The layer number prediction module of the first convolutional neural network processes the first feature map to obtain a first predicted layer number of the sacroiliac joint;

[0073] Calculate a second error value between the first predicted number of layers and the first true number of layers of the sacroiliac joint by using MseLoss;

[0074] L_{Layer1}=(predictive_layer1-true_layer1)^2;

[0075] Where: L_{Layer1} is the second error value; predictive_layer1 is the first prediction layer number;

[0076] true_layer1 is the first true layer number;

[0077] Adding the first error value and the second error value to obtain a third error value;

[0078] L_{Total1}=L_{ROI1}+L_{Layer1}

[0079] The third error value is gradient-backed to adjust the parameters of the first convolutional neural network.

[0080] In the actual application training of the first convolutional neural network, the previous training steps are repeated until the performance of the first convolutional neural network reaches a preset level or meets other stopping conditions.

[0081] In this embodiment, multi-task joint training is designed for the first convolutional neural network, and the first error value between the first predicted target area and the first true target area is calculated by DiceLoss, and the second error value between the first predicted layer and the first true layer of the sacroiliac joint is calculated by MseLoss, and the second error value is used as the model supervision of the first convolutional neural network, which provides a stable gradient direction for the optimization algorithm of the first convolutional neural network and helps to smoothly train the model; and the introduction of the second error value through the MseLoss function can effectively punish the first neural convolutional network for large prediction errors; and in the MRI image segmentation task, the first convolutional neural network will be driven to more accurately predict the pixel label, thereby improving the accuracy of segmentation; and the second error value can be used as auxiliary supervision to provide more fine-grained feedback information, which helps the model to better learn the details of the data.

[0082] See also Figure 2 In one optional implementation, the second convolutional neural network is an encoder-decoder architecture; the training of the second convolutional neural network includes:

[0083] The encoder of the second convolutional neural network processes the input second MRI training image and the first predicted target area to obtain a second feature map;

[0084] The decoder of the second convolutional neural network processes the second feature map to obtain a second predicted target area;

[0085] Calculate a fourth error value between the second predicted target area and the second true target area by using DiceLoss;

[0086] L_{ROI2}=DiceLoss(predictive_ROI2, GroundTruth2);

[0087] Where: L_{ROI2} is the fourth error value; DiceLoss is the loss function; predictive_ROI2 is the second predicted target area; GroundTruth2 is the second true target area;

[0088] The layer number prediction module of the second convolutional neural network processes the second feature map to obtain a second predicted layer number of the sacroiliac joint;

[0089] Calculate a fifth error value between the second predicted number of layers and the second true number of layers of the sacroiliac joint by using MseLoss;

[0090] L_{Layer2}=(predictive_layer2-true_layer2)^2;

[0091] Where: L_{Layer2} is the fifth error value; predictive_layer2 is the second prediction layer number;

[0092] true_layer2 is the second true layer number;

[0093] Adding the fourth error value and the fifth error value to obtain a sixth error value;

[0094] L_{Total2}=L_{ROI2}+L_{Layer2}

[0095] The sixth error value is gradient-backed to adjust the parameters of the second convolutional neural network.

[0096] In this embodiment, multi-task joint training is designed for the second convolutional neural network. The second convolutional neural network and the first convolutional neural network use the same training grid structure. Therefore, the training method of the second convolutional neural network has the same beneficial effects as the first convolutional neural network. The beneficial effects of the second convolutional neural network are not repeated here. For details, please refer to the previous description of the first convolutional neural network.

[0097] In one optional implementation, the layer number prediction module of the first convolutional neural network includes several convolutional layers, pooling layers and linear layers; the layer number prediction module of the second convolutional neural network includes several convolutional layers, pooling layers and linear layers.

[0098] In one optional implementation manner, the performing heterogeneity processing on the input first MRI image includes:

[0099] The size of the first MRI image is adjusted to a preset size, the distribution of voxel intensity values ​​of the first MRI image is truncated within a preset interval, and the truncated voxel intensity values ​​are normalized.

[0100] In this embodiment, the first MRI image is resized to a preset size, which can ensure the consistency of the adjusted image size; at the same time, the distribution of the first MRI is truncated within a preset interval, which can effectively reduce the range of intensity values, thereby solving the problem of highly skewed distribution of voxel intensity values ​​in the first MRI image and reducing the impact of abnormal values ​​on the image; the truncated voxel intensity is normalized to between 0 and 1, which can improve its applicability as a deep neural network input.

[0101] In one specific embodiment, the data of the first MRI image is in DICOM format, and SimpleITK is used to process the first MRI image; it is understandable that the image processing is not limited to the SimpleITK tool, and other tools that can also implement the scheme recorded in this application can also be included in the protection scope, such as ANTs, Elastix and SimpleElastix. Then the first MRI image is adjusted to a 512×512×N matrix size to ensure the size of the image, where N is the number of layers of the first MRI image along the Z axis; in some other embodiments, the first MRI image can also be adjusted to a matrix size of other sizes. Finally, the distribution of the voxel intensity value of the first MRI image is truncated in a preset interval, and the preset interval can be 1% to 99%; the truncated voxel intensity can be normalized by a linear stretching method.

[0102] In one optional implementation, the first convolutional neural network and the second convolutional neural network both adopt a 3D U-Net network.

[0103] A second aspect of the embodiment of the present application provides an MRI image sacroiliac joint segmentation device based on a convolutional neural network, comprising:

[0104] A data preprocessing module, used for performing heterogeneity processing on the input first MRI image to obtain a second MRI image;

[0105] A coarse-grained segmentation module, configured to divide the second MRI image according to the first label by using a first convolutional neural network to obtain a first target area;

[0106] A fine-grained segmentation module is used to divide the second MRI image according to the second label and the first target area through a second convolutional neural network to obtain a second target area; wherein the second target area includes eight quadrants of the sacroiliac joint.

[0107] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0108] An embodiment of the present application further provides a server, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the methods listed above when executing the computer program.

[0109] An embodiment of the present application further provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, the step features included in the above-mentioned systems or devices that can be implemented by the computer program are implemented.

[0110] The processor referred to in this application may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0111] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or circuit that can carry the computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0112] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0113] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, only the division of the above-mentioned functional elements and circuits is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional elements and circuits as needed, that is, the internal structure of the circuit is divided into different functional elements or circuits to complete all or part of the functions described above. The functional elements and circuits in the embodiments can be integrated into a processing element, or each element can exist physically alone, or two or more elements can be integrated into one element. The above-mentioned integrated elements can be implemented in the form of hardware or in the form of software functional elements. In addition, the specific names of the functional elements and circuits are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the elements and circuits in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0114] Those of ordinary skill in the art will appreciate that the elements and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0115] References to "various embodiments", "in an embodiment", "one embodiment", or "an embodiment" throughout the specification, etc., mean that a particular feature, structure, or characteristic described with respect to an embodiment is included in at least one embodiment. Therefore, the appearance of the phrases "in various embodiments", "in some embodiments", "in an embodiment", or "in an embodiment", etc., in appropriate places throughout the specification, do not necessarily refer to the same embodiment. In addition, particular features, structures, or characteristics may be combined in any appropriate manner in one or more embodiments. Therefore, a particular feature, structure, or characteristic shown or described with respect to one embodiment may be combined in whole or in part with features, structures, or characteristics of one or more other embodiments without assuming that such a combination is not an illogical or non-functional limitation. Any directional references (e.g., plus, minus, upper, lower, up, down, left, right, leftward, rightward, top, bottom, above, below, vertical, horizontal, clockwise, and counterclockwise) are used for identification purposes to help the reader understand the present disclosure, and do not create limitations, particularly with respect to the location, orientation, or use of the embodiments.

[0116] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0117] Those of ordinary skill in the art will appreciate that the elements and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0118] In the embodiments provided in the present application, it should be understood that the disclosed circuits / terminal devices and methods can be implemented in other ways. For example, the circuit / terminal device embodiments described above are only schematic. For example, the division of the circuits or components is only a logical function division. There may be other division methods in actual implementation, such as multiple components or circuits can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of circuits or components, which can be electrical, mechanical or other forms.

[0119] The components described as separate components may or may not be physically separated, and the components shown as components may or may not be physical components, that is, they may be located in one place or distributed on multiple network components. Some or all of the components may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0120] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A convolutional neural network-based MRI image sacroiliac joint segmentation method, characterized in that: include: Performing heterogeneity processing on the input first MRI image to obtain a second MRI image; Dividing the second MRI image according to the first label by a first convolutional neural network to obtain a first target area; Dividing the second MRI image according to the second label and the first target area through a second convolutional neural network to obtain a second target area; wherein the second target area includes eight quadrants of the sacroiliac joint; The dividing the second MRI image according to the first label to obtain the first target area includes: Divide the second MRI image according to the first label to obtain the target layer number of the sacroiliac joint and the first target area; The first convolutional neural network is an encoder-decoder architecture; the training of the first convolutional neural network includes: The encoder of the first convolutional neural network processes the input second MRI training image to obtain a first feature map; The decoder of the first convolutional neural network processes the first feature map to obtain a first predicted target area; Calculate a first error value between the first predicted target area and the first real target area by using DiceLoss; L_{ROI1}=DiceLoss(predictive_ROI1, GroundTruth1); Where: L_{ROI1} is the first error value; DiceLoss is the loss function; predictive_ROI1 is the first predicted target area; GroundTruth1 is the first true target area; The layer number prediction module of the first convolutional neural network processes the first feature map to obtain a first predicted layer number of the sacroiliac joint; Calculate a second error value between the first predicted number of layers and the first true number of layers of the sacroiliac joint by using MseLoss; L_{Layer1}=(predictive_layer1-true_layer1)^2; Where: L_{Layer1} is the second error value; predictive_layer1 is the first predicted layer number; true_layer1 is the first true layer number; Adding the first error value and the second error value to obtain a third error value; performing gradient feedback on the third error value to adjust the parameters of the first convolutional neural network; The second convolutional neural network is an encoder-decoder architecture; the training of the second convolutional neural network includes: The encoder of the second convolutional neural network processes the input second MRI training image and the first predicted target area to obtain a second feature map; The decoder of the second convolutional neural network processes the second feature map to obtain a second predicted target area; Calculate a fourth error value between the second predicted target area and the second true target area by using DiceLoss; L_{ROI2}=DiceLoss(predictive_ROI2, GroundTruth2); Where: L_{ROI2} is the fourth error value; DiceLoss is the loss function; predictive_ROI2 is the second predicted target area; GroundTruth2 is the second true target area; The layer number prediction module of the second convolutional neural network processes the second feature map to obtain a second predicted layer number of the sacroiliac joint; Calculate a fifth error value between the second predicted number of layers and the second true number of layers of the sacroiliac joint by using MseLoss; L_{Layer2}=(predictive_layer2-true_layer2)^2; Where: L_{Layer2} is the fifth error value; predictive_layer2 is the second predicted layer number; true_layer2 is the second true layer number; The fourth error value and the fifth error value are added to obtain a sixth error value; and the sixth error value is gradient-transferred back to adjust the parameters of the second convolutional neural network.

2. The method according to claim 1, characterized in that The layer number prediction module of the first convolutional neural network includes several convolutional layers, several pooling layers and several linear layers; the layer number prediction module of the second convolutional neural network includes several convolutional layers, several pooling layers and several linear layers.

3. The method according to any one of claims 1 to 2, characterized in that: The step of performing heterogeneity processing on the input first MRI image comprises: The size of the first MRI image is adjusted to a preset size, the distribution of voxel intensity values ​​of the first MRI image is truncated within a preset interval, and the truncated voxel intensity values ​​are normalized.

4. The method according to any one of claims 1 to 2, characterized in that: The first convolutional neural network and the second convolutional neural network both adopt a 3D U-Net network.

5. A convolutional neural network-based MRI image sacroiliac joint segmentation device, characterized in that: The device applies the method according to any one of claims 1 to 4, including: A data preprocessing module, used for performing heterogeneity processing on the input first MRI image to obtain a second MRI image; A coarse-grained segmentation module, configured to divide the second MRI image according to the first label by using a first convolutional neural network to obtain a first target area; A fine-grained segmentation module is used to divide the second MRI image according to the second label and the first target area through a second convolutional neural network to obtain a second target area; wherein the second target area includes eight quadrants of the sacroiliac joint.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

7. A server comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 4 is implemented.

Images