A brain hippocampus segmentation method and system based on three-dimensional medical images

Abstract

The present invention discloses a brain hippocampus segmentation method and system based on three-dimensional medical images, and relates to the technical field of medical image processing. The method comprises the steps of: acquiring three-dimensional brain medical image volume data; extracting and cutting the three-dimensional hippocampus region of interest, and segmenting it using a segmentation network based on adaptive fusion of spatial domain and frequency domain features, and then further refining the segmentation using inter-class boundary region enhancement technology; mapping the segmentation result of the region of interest to the complete three-dimensional brain medical image volume data, and performing post-processing operations to obtain the final brain hippocampus segmentation result. The present invention solves the problems of difficulty in feature extraction and inaccurate segmentation results caused by various artifacts and noises in brain MRI images, unclear boundaries between the hippocampus as a whole and its surrounding tissues and organs, and less obvious boundaries between the head, body, and tail of the hippocampus, and significant differences in tissues and organs between different subjects.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a brain hippocampus segmentation method and system based on three-dimensional medical images. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Image segmentation refers to dividing an image into multiple specific regions, and it has extensive and important applications in various fields. Medical image segmentation is an important manifestation of this field. It can make the anatomical and pathological structures in the image clearer. It is a preprocessing step for clinical applications such as computer-aided diagnosis, computer-aided surgical planning, clinical quantification and treatment, and smart medical care, and plays an important role in clinical work. However, most medical image interpretations are completed by radiologists, which is not only time-consuming and labor-intensive, but also cannot guarantee its real-time performance; due to the influence of personal subjective factors and work fatigue of doctors, there will be certain differences between the segmentation results; and the doctor's interpretation is limited. The increasing image output and workload will bring negative effects such as missed diagnosis and long work cycle, which greatly restricts the development of the medical field. Therefore, automatic segmentation methods came into being. A complete set of computer-aided medical image segmentation methods and systems can provide support for doctors through automated analysis, which has brought great convenience to clinical work.

[0004] The existing automatic segmentation methods can be divided into semi-automatic segmentation methods and fully automatic segmentation methods according to their automation level, and the inventors have found that according to the different segmentation technologies, they can also be divided into traditional medical image segmentation methods and medical image segmentation methods based on deep learning. These methods contain the creativity of researchers and are very innovative, but when applied to clinical work, they still encounter huge difficulties and challenges.

[0005] Traditional image segmentation methods have achieved good results to a certain extent. At present, some scholars use threshold-based segmentation methods to divide the pixels of an image into several classes using one or more thresholds. Each class is considered to have the same attributes. This method is very simple, but the segmentation effect is directly affected by the threshold, but it is difficult to obtain an optimal threshold; some scholars use edge detection-based segmentation methods to first determine the edge pixels in the image, and then connect the edge pixels to obtain the segmentation result. Usually, some edge detection operators such as Sobel operator and Canny operator are used; some scholars use clustering-based segmentation methods to reveal the rules between data through unlabeled sample training and learning, and obtain multi-category segmentation results by specifying the number of clusters; some scholars use registration-based segmentation methods to register the already segmented atlas image with the target image, and the obtained transformation parameters are applied to the segmentation result of the atlas image to obtain the segmentation result of the target image. There are more methods, including segmentation methods based on variable line models, region-based segmentation methods, active contour-based segmentation methods, machine learning-based segmentation methods, etc. However, since the above methods require manually crafted features, feature representation is difficult and requires various manual intervention methods and certain professional knowledge, which makes the segmentation cost still high. In particular, feature extraction of medical images is more difficult than that of ordinary RGB images, and their images are more blurred, noisy, and have lower contrast.

[0006] In recent years, with the development of deep learning technology, its application has penetrated into all walks of life, and the medical field is no exception. Thanks to the powerful feature extraction ability of convolutional neural networks, it abandons the traditional hand-made features, is not sensitive to image blur, noise and low contrast, and shows superior performance, becoming the first choice in the field of medical image segmentation. The first to emerge is the fully convolutional neural network FCN, which is the most basic encoder-decoder architecture model. The encoder is used to extract image features, and the decoder is used to restore the extracted features to the original image size and output the segmentation results. Inspired by this, the image segmentation method U-Net, which can be widely used in the medical field, was born. It combines low-resolution and high-resolution feature maps through jump connections, effectively integrates low-resolution and high-resolution image features, and perfectly solves the task of medical image segmentation. It is suitable for various medical images such as CT, MRI, X-ray, and ultrasound. The system has 2D U-Net for two-dimensional images and 3D U-Net for three-dimensional volume data, as well as various variants based on this, such as U-Net++, etc., and has become the benchmark and backbone in the field of medical image segmentation.

[0007] Magnetic resonance imaging (MRI) is one of the main imaging methods for medical imaging. It can observe the structural details of the brain, chest and abdomen, liver, pelvis and various soft tissues, but its imaging time is long and there are many influencing factors, which will cause various artifacts and noise in the image: partial volume effect, random field noise, intensity non-uniformity, gradient artifact, motion artifact, surround artifact and magnetic susceptibility artifact, etc.; the tissues and organs of different subjects will also have significant morphological differences; the boundaries of some tissues and organs are low in contrast, especially for the hippocampus of the brain. The boundary between the hippocampus as a whole and its surrounding tissues and organs is not clear, and the direct boundary between the head, body and tail of the hippocampus is even less obvious. All these have led to the fact that using deep learning technology to segment the hippocampus of the brain is still a challenging problem. Summary of the invention

[0008] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a brain hippocampus segmentation method and system based on three-dimensional medical images, which solves the problems of various artifacts and noise in brain MRI images, unclear boundaries between the hippocampus as a whole and its surrounding tissues and organs, and even less obvious boundaries between the head, body, and tail of the hippocampus, and significant differences in tissues and organs between different subjects, resulting in difficult feature extraction and inaccurate segmentation results.

[0009] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0010] The first aspect of the present invention provides a brain hippocampus segmentation method based on three-dimensional medical images, comprising the following steps:

[0011] Acquire three-dimensional brain medical image volume data to be segmented;

[0012] Extracting and cropping a three-dimensional hippocampal region of interest in the three-dimensional brain medical image volume data, and recording the position information of the cropped three-dimensional hippocampal region of interest in the complete volume data;

[0013] Perform data preprocessing and data enhancement operations on the three-dimensional hippocampal region of interest volume data;

[0014] The segmentation network based on adaptive fusion of spatial and frequency domain features was used to segment the 3D hippocampal region of interest after data preprocessing and data enhancement operations, and the preliminary segmentation results were obtained.

[0015] Use the inter-class boundary region enhancement technology to further refine the initial segmentation result to obtain the intermediate segmentation result;

[0016] According to the recorded position information of the 3D hippocampus region of interest in the complete volume data, the segmentation result of the region of interest is mapped to the complete 3D brain medical image volume data, and post-processing operations are performed to obtain the final brain hippocampus segmentation result;

[0017] The final brain hippocampus segmentation results are visualized in three dimensions, and the volume and surface area of ​​each segmentation category and the whole are calculated.

[0018] Further, the specific steps of extracting and cropping the three-dimensional hippocampal region of interest in the three-dimensional brain medical image volume data, and recording the position information of the three-dimensional hippocampal region of interest obtained after cropping in the complete volume data are:

[0019] Determine the bounding box of the hippocampus region based on the annotation file;

[0020] Determine the center point of the hippocampal space based on the bounding box and record the position of the center point of the space;

[0021] The center point is expanded to a set size in the direction of the three-dimensional coordinate axis to obtain the region of interest, and the vertex position information of the region of interest is calculated and recorded according to the position of the spatial center point and the expansion length;

[0022] The region of interest was cropped to obtain three-dimensional hippocampal region of interest volume data.

[0023] Furthermore, the specific steps of performing data preprocessing and data enhancement operations on the three-dimensional hippocampal region of interest volume data are as follows:

[0024] Perform contrast stretching and histogram equalization preprocessing operations on the three-dimensional hippocampal region of interest volume data;

[0025] Calculate the mean and standard deviation of the three-dimensional hippocampal region of interest volume data, and perform Z-Score normalization operation on the three-dimensional hippocampal region of interest volume data according to the calculation results;

[0026] The normalized three-dimensional hippocampal region of interest volume data were augmented using random rotation, random mirroring and random scaling data augmentation methods.

[0027] Furthermore, the segmentation network based on the adaptive fusion of spatial domain and frequency domain features is used to segment the three-dimensional hippocampal region of interest after data preprocessing and data enhancement operations. The specific steps to obtain the preliminary segmentation results are as follows:

[0028] Input the three-dimensional hippocampal region of interest volume data into the backbone network for encoding and feature extraction to obtain initial features;

[0029] The preliminary features are input into the spatial domain and frequency domain feature adaptive fusion module to further enhance the feature extraction capability of the network and obtain fused features;

[0030] The fused features are decoded and the feature map size is restored, and the preliminary segmentation results are obtained through a deep supervision mechanism.

[0031] Furthermore, the specific steps of using the inter-class boundary region enhancement technology to further refine the initial segmentation result to obtain the intermediate segmentation result are as follows:

[0032] The inter-class boundary area map is obtained by using the binary label expansion intersection method, and the inter-class boundary area map is used as the attention map to be learned during each stage of network training;

[0033] The loss function is calculated based on the predicted segmentation results and the true labels, and the weight of the part occupied by the inter-class boundary area map is increased in the loss function.

[0034] Furthermore, according to the recorded position information of the 3D hippocampus ROI in the complete volume data, the segmentation result of the ROI is mapped to the complete 3D brain medical image volume data, and post-processing operation is performed to obtain the final brain hippocampus segmentation result. The specific steps are as follows:

[0035] According to the position information of the recorded three-dimensional hippocampal region of interest in the complete volume data, the segmentation result of the region of interest is mapped back;

[0036] Perform post-processing operations to remove scattered points on the segmentation results of the mapped region of interest;

[0037] Save the complete segmentation result after post-processing operations.

[0038] Furthermore, according to the final brain hippocampus segmentation results in NIFTI format, an obj three-dimensional model file is constructed and an mtl material file is designed to achieve three-dimensional visualization display.

[0039] The second aspect of the present invention provides a brain hippocampus segmentation system based on three-dimensional medical images, comprising:

[0040] A data acquisition module is configured to acquire three-dimensional brain medical image volume data to be segmented; extract and crop a three-dimensional hippocampal region of interest in the three-dimensional brain medical image volume data, and record the position information of the three-dimensional hippocampal region of interest obtained after cropping in the complete volume data; perform data preprocessing and data enhancement operations on the three-dimensional hippocampal region of interest volume data;

[0041] The region segmentation module is configured to segment the three-dimensional hippocampal region of interest after data preprocessing and data enhancement operations using a segmentation network based on adaptive fusion of spatial domain and frequency domain features to obtain a preliminary segmentation result; and further refine the preliminary segmentation result using an inter-class boundary region enhancement technique to obtain an intermediate segmentation result;

[0042] The segmentation optimization module is configured to map the segmentation result of the region of interest to the complete three-dimensional brain medical image volume data according to the position information of the recorded three-dimensional hippocampus region of interest in the complete volume data, and perform post-processing operations to obtain the final brain hippocampus segmentation result;

[0043] The segmentation display module is configured to perform three-dimensional visualization of the final brain hippocampus segmentation results and calculate the volume and surface area of ​​each segmentation category and the whole.

[0044] The third aspect of the present invention provides a medium having a program stored thereon, which, when executed by a processor, implements the steps in the brain hippocampus segmentation method based on three-dimensional medical images as described in the first aspect of the present invention.

[0045] The fourth aspect of the present invention provides a device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the brain hippocampus segmentation method based on three-dimensional medical images as described in the first aspect of the present invention are implemented.

[0046] One or more of the above technical solutions have the following beneficial effects:

[0047] The present invention proposes a brain hippocampus segmentation method and system based on three-dimensional medical images, which uses a spatial domain and frequency domain feature adaptive fusion strategy to enhance the feature extraction capability of the network, greatly reduce the influence of complex artifacts and noise in MRI images on the segmentation effect, and improve the segmentation accuracy of the brain hippocampus.

[0048] The present invention uses an adaptive fusion strategy of spatial domain and frequency domain features to enhance the feature extraction capability of the network, greatly reduce the influence of complex artifacts and noise in MRI images on the segmentation effect, and improve the segmentation accuracy of the brain hippocampus.

[0049] The present invention uses inter-class boundary region enhancement technology to solve the problem that the boundary between the hippocampus as a whole and its surrounding tissues and organs is unclear and the boundary between the head, body and tail inside the hippocampus is even less obvious. It significantly improves the boundary division between the head, body and tail of the hippocampus, ensuring that the volume and surface area calculated for each part are more accurate.

[0050] The present invention can also dynamically display the three-dimensional morphology and the volume and surface area of ​​each part, which can help clinicians to make more accurate disease diagnosis and disease tracking.

[0051] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0053] Figure 1 This is a flow chart of a method for segmenting the brain hippocampus in medical images according to Embodiment 1 of the present invention;

[0054] Figure 2 This is a network model framework diagram of the first embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram of the coding block structure of the first embodiment of the present invention;

[0056] Figure 4 This is a schematic diagram of the structure of a spatial domain and frequency domain adaptive fusion module according to the first embodiment of the present invention;

[0057] Figure 5 This is a schematic diagram of a decoding block structure according to Embodiment 1 of the present invention;

[0058] Figure 6 This is a schematic diagram of the structure of the inter-class boundary region enhancement technology according to the first embodiment of the present invention;

[0059] Figure 7 This is a display diagram of the medical image brain hippocampus visualization system according to the first embodiment of the present invention;

[0060] Figure 8 This is a diagram showing the results of brain hippocampus segmentation in medical images according to the first embodiment of the present invention. DETAILED DESCRIPTION

[0061] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0062] It should be noted that in the embodiments of the present invention, related data such as three-dimensional brain medical imaging volume data are involved. When the above embodiments of the present invention are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards of relevant countries and regions.

[0063] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "include" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or their combinations;

[0064] Embodiment 1:

[0065] Embodiment 1 of the present invention provides a brain hippocampus segmentation method based on three-dimensional medical images, such as Figure 1 As shown, the following steps are included:

[0066] Step 1: Obtain three-dimensional brain medical image volume data to be segmented.

[0067] Step 2: extract and crop the three-dimensional hippocampal region of interest in the three-dimensional brain medical image volume data, and record the position information of the three-dimensional hippocampal region of interest obtained after cropping in the complete volume data.

[0068] Step 3: perform data preprocessing and data enhancement operations on the three-dimensional hippocampal region of interest volume data.

[0069] Step 4: Use a segmentation network based on adaptive fusion of spatial domain and frequency domain features to segment the three-dimensional hippocampal region of interest after data preprocessing and data enhancement operations to obtain a preliminary segmentation result.

[0070] Step 5: Use the inter-class boundary region enhancement technology to further refine the preliminary segmentation result to obtain the intermediate segmentation result.

[0071] Step 6, mapping the segmentation result of the region of interest to the complete three-dimensional brain medical image volume data according to the recorded position information of the three-dimensional hippocampus region of interest in the complete volume data, and performing post-processing operations to obtain the final brain hippocampus segmentation result.

[0072] Step 7: Perform three-dimensional visualization of the final brain hippocampus segmentation results, and calculate the volume and surface area of ​​each segmentation category and the whole.

[0073] In step 1, the specific process of obtaining the three-dimensional brain MRI medical image volume data to be segmented includes:

[0074] Step 1.1: Recruit volunteers and perform head scans using an MRI scanner to obtain brain MRI image data in DICOM format.

[0075] In a specific embodiment, 90 Chinese volunteers were recruited, including 50 males and 40 females, with an average age of 17.30±1.58 years. A 3.0T GE (General Electric, Milwaukee, USA) MRI scanner with a standard eight-channel head coil was used. The parameters of the T1-weighted structural MRI scan were as follows: repetition time (TR) = 18 ms, echo time (TE) = 10 ms, voxel size = 0.47 mm × 0.47 mm × 0.70 mm, field of view (FOV) = 24.0 cm × 24.0 cm, matrix size = 512 × 512, angle = 10°, slice thickness = 1.4 mm, slice gap = 0.7 mm, number of excitations (NEX) = 2, and total scanning time = 12 min. All images were scanned along the horizontal line passing through the anterior-posterior continuum, and brain MRI image data in DICOM format were obtained.

[0076] Step 1.2: Read DICOM format data and convert it to NIFTI format for storage to facilitate subsequent medical image processing.

[0077] Step 1.2.1: First, use the for loop and os.listdir method to traverse the folders where all DICOM format brain MRI image data are located to perform batch processing of the data.

[0078] Step 1.2.2: Use the SimpleITK library to read DICOM format data, use the ImageSeriesReader method to create an image sequence reading object, and use the GetGDCMSeriesFileNames method of the object to obtain the full path of all slices in the DICOM sequence.

[0079] Step 1.2.3: Use the SetFileNames method to read the DICOM sequence through the complete path of all slices in the sequence, and then use the Execute method to obtain the three-dimensional volume data of the sequence.

[0080] Step 1.2.4: Finally, use the WriteImage method to store the 3D volume data as NIFTI format data.

[0081] Step 1.3: Use the load method in the nibabel library to read NIFTI format data into memory, use the get_fdata method to obtain the 3D volume data in the numpy ndarray format array in Python, and use the affine attribute to obtain the affine transformation of the 3D volume data array to the standard reference space.

[0082] In step 2, the specific steps of extracting and cropping the three-dimensional hippocampal region of interest in the three-dimensional brain medical image volume data, and recording the position information of the three-dimensional hippocampal region of interest obtained after cropping in the complete volume data are:

[0083] Step 2.1: Determine the bounding box of the hippocampus region based on the annotation file.

[0084] Step 2.1.1: First, use the ReadImage method in the SimpleITK library to read the three-dimensional brain MRI medical image volume data and label data in NIFTI format, and use the GetArrayFromImage method to convert them into array form.

[0085] Step 2.1.2: Determine the coordinate system based on the three-dimensional array of label data<z,x,y> The minimum non-zero voxel position in the three axis directions<z_min,x_min,y_min> and the maximum non-zero voxel position

[0086] <z_max,x_max,y_max> , and then use six planes passing through these six positions and parallel to the three axes of the coordinate system to draw the rectangular bounding box.

[0087] Step 2.2: Determine the center point of the hippocampal space based on the bounding box and record the position of the center point of the space.

[0088] In this embodiment, the following formula is used to calculate the center point position of the three-dimensional hippocampal region of interest:

[0089] <z_cen,x_cen,y_cen> :

[0090]

[0091] Step 2.3: Expand the center point to the three-dimensional coordinate axis direction (six in total) to the set size to obtain the region of interest, and calculate and record the vertex position information of the region of interest based on the spatial center point position and the expansion length.

[0092] In a specific implementation, the center point is expanded to a specified size in the direction of the three-dimensional coordinate axis (a total of six).<z_len×2,x_len×2,y_len×2> , get the region of interest, and according to the spatial center point position and expansion length<z_len,x_len,y_len> Calculate and record the eight vertex position information of the area of ​​interest <(z_cen-z_len, x_cen-x_len, y_cen-y_len), (z_cen-z_len, x_cen+x_len, y_cen-y_len), (z_cen-z_len, x_cen-x_len, y_cen+y_len), (z_cen-z_len, x_cen-x_len, y_cen+y_len), (z_cen-z_len, x_cen+x_len, y_cen+y_len), (z_cen+

[0093] z_len, x_cen-x_len, y_cen-y_len), (z_cen+z_len, x_cen+x_len, y_cen-y_len), (z_cen+z_len, x_cen-x_len, y_cen+y_len), (z_cen+

[0094] z_len, x_cen+x_len, y_cen+y_len)>.

[0095] Step 2.4: Crop the region of interest to obtain three-dimensional hippocampal region of interest volume data.

[0096] In step 3, the specific steps of performing data preprocessing and data enhancement operations on the three-dimensional hippocampal region of interest volume data are as follows:

[0097] Step 3.1: Perform contrast stretching and histogram equalization preprocessing operations on the three-dimensional hippocampal region of interest volume data.

[0098] Step 3.1.1: Use the cumulative_distribution method in the exposure module to calculate the grayscale distribution histogram of the three-dimensional region of interest volume data.

[0099] Step 3.1.2: Use the rescale_intensity method to intercept the voxel grayscale values ​​in the distribution interval of [2%, 98%]. The grayscale values ​​on both sides are unified and set to the interval boundary grayscale values ​​nearby. This method can eliminate the influence of voxel points with a small number but extremely large or small grayscale values ​​on the contrast of the entire image on both sides. After compressing them to the interval boundary values, the contrast can be stretched to achieve the purpose of enhancing the contrast.

[0100] Step 3.1.3: Then use the equalize_hist method to perform histogram equalization. The voxel values ​​are evenly mapped to a new histogram with uniform grayscale distribution through the mapping function, which can also achieve the purpose of enhancing contrast.

[0101] Step 3.2: Calculate the mean and standard deviation of the three-dimensional hippocampal region of interest volume data, and perform a Z-Score normalization operation on the three-dimensional hippocampal region of interest volume data based on the calculation results.

[0102] In a specific implementation, voxel points with grayscale values ​​greater than 0 in the volume data of the three-dimensional region of interest are counted, their average value voxel_mean and standard deviation voxel_std are calculated, and then the following formula is used to perform Z-Score normalization operation on all voxel points voxel_point greater than 0:

[0103] voxel_point=(voxel_point-voxel_mean) / voxel_std.

[0104] Step 3.3: Use random rotation, random mirroring and random scaling to perform data enhancement operations on the normalized three-dimensional hippocampal region of interest volume data in turn, increase the amount of trained data, prevent network overfitting, and improve the robustness and generalization ability of the network.

[0105] In step 4, the segmentation network based on the adaptive fusion of spatial and frequency domain features is used to segment the 3D hippocampal region of interest after data preprocessing and data enhancement operations. The segmentation network consists of three parts: a backbone network for encoding and feature extraction, a spatial and frequency domain feature adaptive fusion module, and a decoding module with a deep supervision mechanism. Figure 2 As shown in the designed brain hippocampus segmentation network framework, the specific steps to obtain the preliminary segmentation results are:

[0106] Step 4.1: Input the three-dimensional hippocampal region of interest volume data into the backbone network for encoding and feature extraction to obtain initial features.

[0107] In a specific implementation, the feature extraction of the backbone network mainly consists of four coding blocks, each of which is as follows: Figure 3 As shown in the figure, it consists of three convolution blocks. The second convolution block is followed by a random dropout function. Each convolution block consists of a convolution function, a normalization function, and an activation function. After each convolution block, the size of the feature map is reduced to half of the original size, while the number of channels is increased to twice the original size. This can expand the receptive field of the network and extract high-level semantic information at the same time.

[0108] Step 4.2: Input the preliminary features into the spatial domain and frequency domain feature adaptive fusion module to further enhance the feature extraction capability of the network and obtain the fused features.

[0109] In a specific embodiment, Figure 4 As shown in Figure 1, the spatial domain and frequency domain feature adaptive fusion module consists of three branches:

[0110] The first branch is the original feature map path branch. The residual connection can be used to break the symmetry of the network and improve the representation ability of the network. The feature map obtained is F i .

[0111] The second branch is the spatial domain feature extraction branch, which is used to extract all the features contained in the MRI image. It contains two convolution functions and one activation function. The feature map obtained is F s .

[0112] The third branch is the frequency domain feature extraction path branch. Through the convolution operation in the frequency domain, it simulates the waveform combination selection process, filters out various artifacts and noises in the brain MRI images and the features generated by individual differences in the encoding stage, eliminates the negative impact of these features on the segmentation results, and enables the final extracted features to more accurately express the contextual information of the hippocampal head, body, tail and its surroundings.

[0113] First, the feature map extracted by the backbone network coding block is converted from the spatial domain to the frequency domain using a three-dimensional fast Fourier transform to obtain the real part map Freq real and imaginary part graph Freq imag :

[0114] Freqreal,Freqima g =FFT(Fi).

[0115] Among them, FFT(·) represents three-dimensional fast Fourier transform, the real part graph represents the amplitude graph, and the imaginary part graph represents the phase graph. The amplitude and phase can constitute a waveform graph. Since all waves can be composed of infinitely superimposed sine waves, this embodiment selects a suitable waveform combination in the frequency domain graph through convolution to achieve the purpose of filtering invalid features. Specifically:

[0116] Step 4.2.1: Concatenate the real part graph and the imaginary part graph, filter the frequency by combining two convolution functions and an activation function, and obtain the appropriate frequency combination Freq choi :

[0117] Freqchoi=Conv(Activate(Conv(Concat(Freqreal,Freqima g )))).

[0118] Among them, Conv(·) represents the convolution operation, Activate(·) represents the activation function, and Concat(·) represents the concatenation operation.

[0119] Step 4.2.2: Then decompose it into real and imaginary parts again and reconstruct it into complex form, and then convert it to the spatial domain through 3D inverse fast Fourier transform to obtain the feature map F f :

[0120] F f =IFFT(Complex(Chunk(Freq choi ))).

[0121] Among them, IFFT(·) represents the three-dimensional inverse fast Fourier transform, Complex(·) represents the reconstruction of the complex form, and Chunk(·) represents the decomposition operation.

[0122] Step 4.2.3: Finally, use a learnable parameter α with an initial value of 0.5 and continuously adjust it during the network training process to adaptively fuse the feature maps of the three branches to obtain the final extracted feature map F o :

[0123] F o =α·F s +(1-α)·F f +F i .

[0124] The use of the spatial domain and frequency domain feature adaptive fusion module improves the network's robustness to image status and feature extraction capabilities, and solves the problem of inaccurate brain hippocampal segmentation results caused by the presence of various artifacts and noise in brain MRI images and significant differences in tissues and organs between different subjects.

[0125] Step 4.3: Decode the fused features and restore the feature map size, and obtain the preliminary segmentation results through the deep supervision mechanism.

[0126] In a specific implementation, the decoding stage also includes four decoding blocks, each decoding block is as follows: Figure 5 As shown, it is also composed of three convolution blocks, and the second convolution block is followed by an upsampling function to restore the size of the feature map. At the same time, the decoder path receives the encoder feature map transmitted from the jump connection, and obtains more context information at the corresponding position of the corresponding encoder. In addition, this embodiment also uses the idea of ​​the deep supervision mechanism to combine the multi-scale feature maps generated by the 2nd, 3rd and 4th layers in the decoder through weighted summation to obtain the predicted segmentation result.

[0127] In step 5, if Figure 2As shown in the figure, the specific steps of using the inter-class boundary area enhancement technology to further refine the initial segmentation result and obtain the intermediate segmentation result are:

[0128] Step 5.1: Use the binary label expansion intersection method to obtain the inter-class boundary area map, and use the inter-class boundary area map as the attention map to be focused on during each stage of network training.

[0129] Step 5.1.1: Figure 6 As shown, firstly, the binary label images of the hippocampal head, body, tail and background are obtained respectively (Y head ,Y body ,Y tail ,Y bg ), that is, the voxel point corresponding to each class is 1, and other positions are 0.

[0130] Step 5.1.2: Then use a 2×2×2 dilation kernel to dilate each binary label image to expand its volume:

[0131] (Y dil_head ,Y dil_body ,Y dil_tail ,Y dil_bg )=Dilate(Y head ,Y body ,Y tail ,Y bg ).

[0132] Among them, (Y dil_head ,Y dil_body ,Y dil_tail ,Y dil_bg ) is a binary label image of the hippocampus head, body, tail and background after expansion, so that there will be overlapping areas between every two binary label images after the volume expansion, and these overlapping areas are obtained by intersection operation, and there are a total of six in this embodiment.

[0133] Step 5.1.3: Then, these overlapping areas are integrated into a graph through the union operation. These overlapping areas contain the boundaries between the classes and the contextual information around the boundaries. The integrated graph is the inter-class boundary region graph CBR, which will also be the attention graph that the network focuses on learning during each stage of training:

[0134] CBR=∪(∩(Y p ,Y q )).

[0135] Among them, (Y p ,Y q )∈(Ydil_head,Ydil_bod y ,Ydil_tail,Ydil_b g ) and Yp ! =Y q .

[0136] Step 5.2: Calculate the loss function based on the predicted segmentation results and the true label, and increase the weight of the part of the inter-class boundary area map in the loss function.

[0137] In a specific embodiment, the predicted segmentation result and the true label are compared to calculate the loss function, and the extracted boundary region map is used to calculate the multi-class cross entropy loss function at the voxel level. To perform weighting:

[0138]

[0139] in, is the weighted voxel-level multi-class cross entropy loss function, λ is the weight factor, is defined as:

[0140]

[0141] Among them, Y is the true label, For multi-class segmentation prediction, the final loss function of the entire network is is defined as:

[0142]

[0143] in, is the Dice loss function, which is defined as:

[0144]

[0145] Here, |·| represents the number of elements in the set.

[0146] The use of inter-class boundary region enhancement technology improves the network's ability to learn the edge areas of tissues and organs in brain MRI images, solving the problem of inaccurate hippocampus segmentation results caused by unclear boundaries between the hippocampus as a whole and its surrounding tissues and organs, and even less obvious boundaries between the head, body, and tail of the hippocampus.

[0147] In step 6, the segmentation result of the region of interest is mapped to the complete three-dimensional brain medical image volume data according to the position information of the recorded three-dimensional hippocampus region of interest in the complete volume data, and post-processing operations are performed to obtain the final brain hippocampus segmentation result. The specific steps are:

[0148] Step 6.1: Based on the recorded position information of the three-dimensional hippocampal region of interest in the complete volume data, the segmentation result of the region of interest is mapped back.

[0149] Specifically, the eight vertex position information of the recorded region of interest in the complete three-dimensional brain medical image volume data is used, and the voxel value of the segmentation result is used at the corresponding position to overwrite the voxel value of the complete three-dimensional brain medical image volume data to complete the mapping restoration.

[0150] Step 6.2: Perform post-processing on the segmentation results of the mapped region of interest to remove scattered points.

[0151] In a specific implementation, the regionprops method in the skimage.measure module is used to calculate all connected domains of the segmentation results of the region of interest and sort them according to the size of the connected domains. The three largest connected domains are retained, representing the segmentation results of the hippocampal head, body, and tail, respectively. The other connected domains are scattered points and are removed by zeroing, and the post-processing operation is completed.

[0152] Step 6.3: Save the complete segmentation result after post-processing operations.

[0153] In a specific implementation, the GetImageFromArray method in the SimpleITK library is used to obtain volume data from a three-dimensional array, and the WriteImage method is used to save the volume data of the segmentation result in the NIFTI format.

[0154] In step 7, the obj three-dimensional model file is constructed and the mtl material file is designed according to the final brain hippocampus segmentation result in NIFTI format, so as to realize three-dimensional visualization display.

[0155] Step 7.1: Use the Marching Cubes algorithm to build the obj 3D model file based on the segmentation results in NIFTI format.

[0156] Step 7.2: Calculate the volume and surface area of ​​each segmentation category and the whole, and save them.

[0157] Step 7.3: Design the mtl material file for 3D visualization. The result is as follows: Figure 7 shown.

[0158] The present invention provides a method and system for segmenting the brain hippocampus of medical images, which uses a spatial domain and frequency domain feature adaptive fusion module to improve the robustness and feature extraction ability of the network to the image state, solves the problem that the imaging time is long and there are many influencing factors, which will cause the image to contain various artifacts and noises: partial volume effect, random field noise, intensity non-uniformity, gradient artifact, motion artifact, surround artifact and magnetic susceptibility artifact, etc., and the tissues and organs of different subjects will also have significant morphological differences; using the inter-class boundary area enhancement technology, the network's learning ability for the edge area of ​​tissues and organs in brain MRI images is improved, and the low contrast of the boundaries of some tissues and organs is solved, especially for the brain hippocampus, the boundary between the hippocampus as a whole and the surrounding tissues and organs is unclear and the direct boundary of the head, body and tail of the hippocampus is less obvious. Accurate segmentation of the brain hippocampus and the head, body and tail of the hippocampus is achieved.

[0159] The present invention verifies the segmentation effect of the above method on the collected brain MRI medical image data set. The data set is divided into a training set and a test set in a ratio of 4:1. The segmentation results of the hippocampal head, body, and tail are shown in Figure 8 .

[0160] The present invention adopts the Dice index and the Hausdorff_95distance (95% HD) index to quantitatively measure the segmentation effect. The higher the Dice value, the better, and the lower the HD value, the better. Compared with the classic medical segmentation network 3D U-Net, the present invention significantly improves the segmentation accuracy of the hippocampal head, body, and tail, as shown in Table 1:

[0161] Table 1 Experimental results

[0162]

[0163]

[0164] It can be seen from the above table that the segmentation results obtained by the method of the present invention have achieved significant improvements in three indicators: Dice similarity coefficient, Hausdorff 95% distance and average surface distance.

[0165] Embodiment 2:

[0166] Embodiment 2 of the present invention provides a brain hippocampus segmentation system based on three-dimensional medical images, including:

[0167] A data acquisition module is configured to acquire three-dimensional brain medical image volume data to be segmented; extract and crop a three-dimensional hippocampal region of interest in the three-dimensional brain medical image volume data, and record the position information of the three-dimensional hippocampal region of interest obtained after cropping in the complete volume data; perform data preprocessing and data enhancement operations on the three-dimensional hippocampal region of interest volume data;

[0168] The region segmentation module is configured to segment the three-dimensional hippocampal region of interest after data preprocessing and data enhancement operations using a segmentation network based on adaptive fusion of spatial domain and frequency domain features to obtain a preliminary segmentation result; and further refine the preliminary segmentation result using an inter-class boundary region enhancement technique to obtain an intermediate segmentation result;

[0169] The segmentation optimization module is configured to map the segmentation result of the region of interest to the complete three-dimensional brain medical image volume data according to the position information of the recorded three-dimensional hippocampus region of interest in the complete volume data, and perform post-processing operations to obtain the final brain hippocampus segmentation result;

[0170] The segmentation display module is configured to perform three-dimensional visualization of the final brain hippocampus segmentation results and calculate the volume and surface area of ​​each segmentation category and the whole.

[0171] Embodiment three:

[0172] Embodiment 3 of the present invention provides a medium on which a program is stored. When the program is executed by a processor, the steps in the brain hippocampus segmentation method based on three-dimensional medical images as described in Embodiment 1 of the present invention are implemented.

[0173] Embodiment 4:

[0174] Embodiment 4 of the present invention provides a device, including a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in the brain hippocampus segmentation method based on three-dimensional medical images as described in Embodiment 1 of the present invention are implemented.

[0175] The steps involved in the above embodiments 2, 3 and 4 correspond to the method embodiment 1. For the specific implementation methods, please refer to the relevant description part of embodiment 1.

[0176] Those skilled in the art should understand that the modules or steps of the present invention described above can be implemented by a general-purpose computer device, or alternatively, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0177] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A brain hippocampus segmentation method based on three-dimensional medical images, characterized in that: The following steps are involved: Acquire three-dimensional brain medical image volume data to be segmented; Extracting and cropping a three-dimensional hippocampal region of interest in the three-dimensional brain medical image volume data, and recording the position information of the cropped three-dimensional hippocampal region of interest in the complete volume data; Perform data preprocessing and data enhancement operations on the three-dimensional hippocampal region of interest volume data; The segmentation network based on adaptive fusion of spatial and frequency domain features is used to segment the three-dimensional hippocampal region of interest after data preprocessing and data enhancement operations to obtain preliminary segmentation results; the specific steps are as follows: Inputting the three-dimensional hippocampal region of interest volume data into the backbone network for encoding and feature extraction to obtain initial features; the feature extraction of the backbone network consists of four encoding blocks, each encoding block consists of three convolution blocks, the second convolution block is followed by a random inactivation function, and each convolution block consists of a convolution function, a normalization function and an activation function; after each convolution block, the size of the feature map is reduced to half of the original size, and the number of channels is increased to twice the original size; The preliminary features are input into the spatial domain and frequency domain feature adaptive fusion module to further enhance the feature extraction capability of the network and obtain fused features; The spatial domain and frequency domain feature adaptive fusion module consists of three branches: the first branch is the original feature map path branch, which uses residual connections to break the symmetry of the network and improve the representation ability of the network. The feature map obtained is F i ; The second branch is the spatial domain feature extraction branch, which is used to extract all the features contained in the MRI image. It contains two convolution functions and one activation function. The feature map obtained is F s ; The third branch is the frequency domain feature extraction path branch, which simulates the selection process of waveform combinations through convolution operations in the frequency domain, and filters out various artifacts and noises in brain MRI images as well as the features generated by individual differences in the encoding stage; In the process of obtaining fusion features, the feature map extracted by the backbone network coding block is first converted from the spatial domain to the frequency domain using a three-dimensional fast Fourier transform to obtain the real part map Freq real and imaginary part graph Freq imag : Freq real ,Freq imag =FFT(F i ); Among them, FFT(·) represents three-dimensional fast Fourier transform, the real part graph represents the amplitude graph, and the imaginary part graph represents the phase graph. The amplitude and phase can form a waveform graph; The real part graph and the imaginary part graph are spliced ​​together, and the frequency is filtered through a combination of two convolution functions and an activation function to obtain a suitable frequency combination Freq choi : Freq choi =Conv(Activate(Conv(Concat(Freq real ,Freq imag )))); Among them, Conv(·) represents the convolution operation, Activate(·) represents the activation function, and Concat(·) represents the concatenation operation; Then, the feature map extracted by the backbone network coding block is decomposed into real and imaginary parts and reconstructed into complex form, and then converted to the spatial domain through three-dimensional inverse fast Fourier transform to obtain the feature map F f : F f =IFFT(Complex(Chunk(Freq choi ))); Among them, IFFT(·) represents the three-dimensional inverse fast Fourier transform, Complex(·) represents the reconstruction of complex form, and Chunk(·) represents the decomposition operation; By setting an initial value to α and continuously adjusting the learnable parameter α during the network training process, the feature maps of the three branches are adaptively fused to obtain the final extracted feature map F o : F o =α·F s +(1-α)·F f +F i ; Decode the fused features and restore the feature map size, and obtain the preliminary segmentation results through the deep supervision mechanism; Use the inter-class boundary region enhancement technology to further refine the initial segmentation result to obtain the intermediate segmentation result; the specific steps are: The inter-class boundary area map is obtained by using the binary label expansion intersection method, and the inter-class boundary area map is used as the attention map to be learned during each stage of network training; Specifically, the binary label images of the hippocampal head, body, tail and background are obtained respectively (Y head ,Y body ,Y tail ,Y bg ), the voxel point corresponding to each class is 1, and other positions are 0; Use a 2×2×2 dilation kernel to dilate each binary label image to expand its volume: (AND dil_head ,AND dil_body ,AND dil_tail ,AND dil_bg )=Dilate(Y head ,AND body ,AND tail ,AND bg ); Among them, (Y dil_head ,Y dil_body ,Y dil_tail ,Y dil_bg ) is the binary label image of the hippocampus head, body, tail and background after expansion, so that there will be overlapping areas between every two binary label images after the volume expansion, and these overlapping areas are obtained by intersection operation; The overlapping areas are then integrated into a graph through a union operation. These overlapping areas contain the boundaries between the classes and the contextual information around the boundaries. The integrated graph is the inter-class boundary region graph CBR, which is also the attention graph that the network focuses on learning during each stage of training: CBR=∪(∩(Y p ,Y q )); where (Y p , Y q ) ∈ (Y dil_head , Y dil_body , Y dil_tail , Y dil_bg ) and Y p != Y q ; The loss function is calculated based on the predicted segmentation results and the true labels, and the weight of the part of the inter-class boundary area map is increased in the loss function; specifically, the loss function is calculated based on the predicted segmentation results and the true labels, and the extracted boundary area map is used to perform a multi-class cross entropy loss function at the voxel level. To perform weighting: in, is the weighted voxel-level multi-class cross entropy loss function, λ is the weight factor, is defined as: Among them, Y is the true label, For multi-class segmentation prediction, the final loss function of the entire network is is defined as: in, is the Dice loss function, which is defined as: Among them, |·| represents the number of elements in the set; According to the recorded position information of the 3D hippocampus region of interest in the complete volume data, the segmentation result of the region of interest is mapped to the complete 3D brain medical image volume data, and post-processing operations are performed to obtain the final brain hippocampus segmentation result; The final brain hippocampus segmentation results are visualized in three dimensions, and the volume and surface area of ​​each segmentation category and the whole are calculated.

2. The brain hippocampus segmentation method based on three-dimensional medical images as claimed in claim 1, characterized in that: The specific steps of extracting and cropping the three-dimensional hippocampal region of interest in the three-dimensional brain medical image volume data, and recording the position information of the three-dimensional hippocampal region of interest obtained after cropping in the complete volume data are as follows: Determine the bounding box of the hippocampus region based on the annotation file; Determine the center point of the hippocampal space based on the bounding box and record the position of the center point of the space; The center point is expanded to a set size in the direction of the three-dimensional coordinate axis to obtain the region of interest, and the vertex position information of the region of interest is calculated and recorded according to the position of the spatial center point and the expansion length; The region of interest was cropped to obtain three-dimensional hippocampal region of interest volume data.

3. The brain hippocampus segmentation method based on three-dimensional medical images as claimed in claim 1, characterized in that: The specific steps for data preprocessing and data enhancement operations on the three-dimensional hippocampal region of interest volume data are as follows: Perform contrast stretching and histogram equalization preprocessing operations on the three-dimensional hippocampal region of interest volume data; Calculate the mean and standard deviation of the three-dimensional hippocampal region of interest volume data, and perform Z-Score normalization operation on the three-dimensional hippocampal region of interest volume data according to the calculation results; The normalized three-dimensional hippocampal region of interest volume data were augmented using random rotation, random mirroring and random scaling data augmentation methods.

4. The brain hippocampus segmentation method based on three-dimensional medical images as claimed in claim 1, characterized in that: According to the recorded position information of the 3D hippocampus region of interest in the complete volume data, the segmentation result of the region of interest is mapped to the complete 3D brain medical image volume data, and post-processing operations are performed to obtain the final brain hippocampus segmentation result. The specific steps are as follows: According to the position information of the recorded three-dimensional hippocampal region of interest in the complete volume data, the segmentation result of the region of interest is mapped back; Perform post-processing operations to remove scattered points on the segmentation results of the mapped region of interest; Save the complete segmentation result after post-processing operations.

5. The brain hippocampus segmentation method based on three-dimensional medical images as claimed in claim 1, characterized in that: According to the final brain hippocampus segmentation results in NIFTI format, the obj three-dimensional model file is constructed and the mtl material file is designed to achieve three-dimensional visualization display.

6. A brain hippocampus segmentation system based on three-dimensional medical images, characterized in that: include: A data acquisition module is configured to acquire three-dimensional brain medical image volume data to be segmented; extract and crop a three-dimensional hippocampal region of interest in the three-dimensional brain medical image volume data, and record the position information of the three-dimensional hippocampal region of interest obtained after cropping in the complete volume data; Perform data preprocessing and data enhancement operations on the three-dimensional hippocampal region of interest volume data; The region segmentation module is configured to segment the three-dimensional hippocampal region of interest after data preprocessing and data enhancement operations using a segmentation network based on adaptive fusion of spatial domain and frequency domain features to obtain a preliminary segmentation result; the specific steps are: Inputting the three-dimensional hippocampal region of interest volume data into the backbone network for encoding and feature extraction to obtain initial features; the feature extraction of the backbone network consists of four encoding blocks, each encoding block consists of three convolution blocks, the second convolution block is followed by a random inactivation function, and each convolution block consists of a convolution function, a normalization function and an activation function; after each convolution block, the size of the feature map is reduced to half of the original size, and the number of channels is increased to twice the original size; The preliminary features are input into the spatial domain and frequency domain feature adaptive fusion module to further enhance the feature extraction capability of the network and obtain fused features; The spatial domain and frequency domain feature adaptive fusion module consists of three branches: the first branch is the original feature map path branch, which uses residual connections to break the symmetry of the network and improve the representation ability of the network. The feature map obtained is F i ; The second branch is the spatial domain feature extraction branch, which is used to extract all the features contained in the MRI image. It contains two convolution functions and one activation function. The feature map obtained is F s ; The third branch is the frequency domain feature extraction path branch, which simulates the selection process of waveform combinations through convolution operations in the frequency domain, and filters out various artifacts and noises in brain MRI images as well as the features generated by individual differences in the encoding stage; In the process of obtaining fusion features, the feature map extracted by the backbone network coding block is first converted from the spatial domain to the frequency domain using a three-dimensional fast Fourier transform to obtain the real part map Freq real and imaginary part graph Freq imag : Freq real ,Freq imag =FFT(F i ); Among them, FFT(·) represents three-dimensional fast Fourier transform, the real part graph represents the amplitude graph, and the imaginary part graph represents the phase graph. The amplitude and phase can form a waveform graph; The real part graph and the imaginary part graph are spliced ​​together, and the frequency is filtered through a combination of two convolution functions and an activation function to obtain a suitable frequency combination Freq choi : Freq choi =Conv(Activate(Conv(Concat(Freq real ,Freq imag )))); Among them, Conv(·) represents the convolution operation, Activate(·) represents the activation function, and Concat(·) represents the concatenation operation; Then, the feature map extracted by the backbone network coding block is decomposed into real and imaginary parts and reconstructed into complex form, and then converted to the spatial domain through three-dimensional inverse fast Fourier transform to obtain the feature map F f : F f =IFFT(Complex(Chunk(Freq choi ))); Among them, IFFT(·) represents the three-dimensional inverse fast Fourier transform, Complex(·) represents the reconstruction of complex form, and Chunk(·) represents the decomposition operation; By setting an initial value to α and continuously adjusting the learnable parameter α during the network training process, the feature maps of the three branches are adaptively fused to obtain the final extracted feature map F o : F o =α·F s +(1-α)·F f +F i ; Decode the fused features and restore the feature map size, and obtain the preliminary segmentation results through the deep supervision mechanism; Use the inter-class boundary region enhancement technology to further refine the initial segmentation result to obtain the intermediate segmentation result; the specific steps are: The inter-class boundary area map is obtained by using the binary label expansion intersection method, and the inter-class boundary area map is used as the attention map to be learned during each stage of network training; Specifically, the binary label images of the hippocampal head, body, tail and background are obtained respectively (Y head ,Y body ,Y tail ,Y bg ), the voxel point corresponding to each class is 1, and other positions are 0; Use a 2×2×2 dilation kernel to dilate each binary label image to expand its volume: (AND dil_head ,AND dil_body ,AND dil_tail ,AND dil_bg )=Dilate(Y head ,AND body ,AND tail ,AND bg ); Among them, (Y dil_head ,Y dil_body ,Y dil_tail ,Y dil_bg ) is the binary label image of the hippocampus head, body, tail and background after expansion, so that there will be overlapping areas between every two binary label images after the volume expansion, and these overlapping areas are obtained by intersection operation; The overlapping areas are then integrated into a graph through a union operation. These overlapping areas contain the boundaries between the classes and the contextual information around the boundaries. The integrated graph is the inter-class boundary region graph CBR, which is also the attention graph that the network focuses on learning during each stage of training: CBR=∪(∩(Y p ,Y q )); Among them, (Y p , Y q ) ∈ (Y dil_head , Y dil_body , Y dil_tail , Y dil_bg ) and Y p ! = Y q ; The loss function is calculated based on the predicted segmentation results and the true labels, and the weight of the part of the inter-class boundary area map is increased in the loss function; specifically, the loss function is calculated based on the predicted segmentation results and the true labels, and the extracted boundary area map is used to perform a multi-class cross entropy loss function at the voxel level. To perform weighting: in, is the weighted voxel-level multi-class cross entropy loss function, λ is the weight factor, is defined as: Among them, Y is the true label, For multi-class segmentation prediction, the final loss function of the entire network is is defined as: in, is the Dice loss function, which is defined as: Among them, |·| represents the number of elements in the set; The segmentation optimization module is configured to map the segmentation result of the region of interest to the complete three-dimensional brain medical image volume data according to the position information of the recorded three-dimensional hippocampus region of interest in the complete volume data, and perform post-processing operations to obtain the final brain hippocampus segmentation result; The segmentation display module is configured to perform three-dimensional visualization of the final brain hippocampus segmentation results and calculate the volume and surface area of ​​each segmentation category and the whole.

7. A computer-readable storage medium, characterized in that: A plurality of instructions are stored therein, and the instructions are suitable for being loaded by a processor of a terminal device and executing the brain hippocampus segmentation method based on three-dimensional medical images as described in any one of claims 1-5.

8. A terminal device, characterized in that: It includes a processor and a computer-readable storage medium, the processor is used to implement each instruction; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded by the processor and executing the brain hippocampus segmentation method based on three-dimensional medical images described in any one of claims 1-5.

Images