Gynecological tumor MRI image three-dimensional reconstruction method and system based on deep learning

Abstract

This invention relates to the field of image processing technology, and particularly to a method and system for three-dimensional reconstruction of gynecological tumor MRI images based on deep learning. The method includes: processing coarsely labeled tumor regions using a target detection model based on the original image to generate bounding boxes for small lesions; segmenting the images using these bounding boxes to obtain images of small and non-small lesions; proportionally reducing the size of the small lesions to obtain lesion images; selecting samples based on the generated boundary regions, calculating the mean grayscale value and difference coefficient, and weighted merging to obtain normal tissue images; compressing the tumor core region and boundary region to obtain the overall tumor image; stitching the lesion images, normal tissue images, and tumor images together to form a low-resolution image to train the model, and then segmenting and extracting features to construct a three-dimensional tumor model. This invention preserves the features of small lesions and boundaries, improves segmentation accuracy, constructs a high-precision three-dimensional model, and adapts to clinical diagnostic needs.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for three-dimensional reconstruction of gynecological tumor MRI images based on deep learning. Background Technology

[0002] With the development of medical technology, MRI imaging technology has been widely used in the field of medical imaging. In current medical imaging, for gynecological tumor MRI data, low resolution can highlight the overall outline of the tumor, while high resolution can highlight the local details of the tumor.

[0003] In existing technologies, to improve the accuracy of detail restoration in tumor MRI image reconstruction and the model's adaptability to low resolution, deep learning is used iteratively from low to high resolution, ultimately enabling the model to stably and accurately identify lesions in clinical applications. However, small lesions in tumors, including tiny depressions or protrusions at the tumor boundary, are swallowed up by downsampling (dimensionality reduction) at low resolution, where multiple adjacent pixels are merged into one pixel. This prevents the model from learning these features during training. Furthermore, due to the difference in grayscale values ​​between the tumor boundary and normal tissue, when pixels from the tumor boundary and normal tissue each occupy a portion and are uniformly merged in the same pixel merging region, the grayscale values ​​of the boundary region are averaged, making it impossible to identify the tumor boundary. During the high-resolution learning stage, the loss of detail caused by low-resolution downsampling makes it impossible to correct global structural biases of the tumor during high-resolution training. Consequently, the trained model cannot accurately identify lesions, ultimately resulting in inaccurate 3D tumor MRI images.

[0004] Therefore, it is necessary to propose a method and system for three-dimensional reconstruction of gynecological tumor MRI images based on deep learning to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for three-dimensional reconstruction of gynecological tumor MRI images based on deep learning, so as to solve the problems of multiple adjacent pixels merging into one pixel, causing the small lesion features on the tumor to be swallowed up, and in the same pixel merging area, when the pixels of the tumor boundary and normal tissue each occupy a part and are uniformly merged, the gray value of the boundary area of ​​the two is averaged.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A deep learning-based method for three-dimensional reconstruction of gynecological tumor MRI images, the method comprising:

[0008] Based on the high-resolution original image, the tumor region marked with coarse labels in the original image is processed using a target detection model to generate a localization box, wherein the localization box is used to identify and locate small lesions.

[0009] Based on the localization results of the bounding box recognition, the original image is segmented to obtain images of small lesions and non-small lesions. The image size of the small lesion image is reduced proportionally to obtain the lesion image. The non-small lesion image includes tumors and normal tissue.

[0010] Based on the original image, the grayscale values ​​of the image are obtained, and a grayscale feature map is generated after fusion. Based on the grayscale gradient of the grayscale feature map, boundary candidate points are generated. The boundary candidate points are clustered in three-dimensional space to generate boundary regions. The minimum bounding cube of each boundary region is calculated to obtain the boundary region, wherein the boundary region includes the tumor boundary and the adjacent normal tissue boundary.

[0011] Samples from the core area of ​​the tumor and samples from normal tissue far from the boundary area were selected, and the mean gray value was calculated for each. Based on the mean gray value, the gray value difference coefficient between the tumor and normal tissue was calculated. Based on the gray value difference coefficient and weighted according to the distance between the normal tissue and the boundary area, the normal tissue image was obtained after pixel merging.

[0012] The tumor core and boundary regions are compressed according to the scaling ratio to obtain the tumor image;

[0013] The lesion image, tumor image, and normal tissue image are stitched together according to their original size ratio and spatial coordinates to obtain a low-resolution image. The low-resolution image is used to train a model, and the trained model is used to segment and extract features from the MRI image to construct a three-dimensional model of the tumor.

[0014] Preferably, the process of segmenting the original image based on the positioning result of the bounding box recognition specifically includes:

[0015] The three-dimensional pixel coordinates of the positioning box are mapped to the original image to obtain the small lesion area in the original image;

[0016] Generate a mask image of the same size as the original image, and mark the small lesion area corresponding to the positioning box and the area outside the positioning box respectively;

[0017] Based on the mask image, the marked small lesion regions in the original image are cropped to generate small lesion images;

[0018] Based on the mask image, the region outside the marked location box in the original image is cropped to generate an image of non-micro lesions.

[0019] Preferably, the process of generating boundary candidate points based on the gray-level gradient of the gray-level feature map specifically includes:

[0020] Based on the grayscale feature map, calculate the three-dimensional grayscale gradient of each pixel, and use the pixels with non-zero gradient values ​​to generate initial boundary candidate points.

[0021] Based on the gray-scale difference coefficient, a dynamic threshold is set, and the initial boundary candidate points are filtered by comparing the gradient value with the dynamic threshold.

[0022] Based on the filtered candidate points, neighborhood analysis is used to obtain consecutive pixels in the neighborhood that have at least two candidate points of the same type.

[0023] Based on the tumor region marked by the coarse label, the boundary between the tumor region and normal tissue is used as a limiting condition to filter continuous pixels in the neighborhood to obtain boundary candidate points.

[0024] Preferably, the process of calculating the minimum circumscribed cuboid of each boundary region to obtain the boundary region specifically includes:

[0025] Obtain the 3D coordinates of all pixels within each boundary region after clustering in 3D space;

[0026] Calculate the maximum and minimum values ​​on the x, y, and z axes based on the three-dimensional coordinates of each pixel in the boundary region;

[0027] The starting vertex is generated by finding the minimum values ​​of the x, y, and z axes of each boundary region, and the diagonal vertex is generated by finding the maximum values ​​of the x, y, and z axes of each boundary region. Based on the starting vertex and the diagonal vertex, the minimum circumscribed cuboid is constructed.

[0028] The boundary region is obtained by determining the spatial extent of the smallest circumscribed cuboid corresponding to each boundary region.

[0029] Preferably, the process of obtaining a normal tissue image by merging pixels based on a grayscale difference coefficient and weighted according to the distance between normal tissue and the boundary region specifically includes:

[0030] Calculate the three-dimensional spatial distance between each pixel and the nearest boundary region for each pixel within the normal tissue region;

[0031] Pixel weights are assigned to normal tissue regions based on grayscale difference coefficients and three-dimensional spatial distance.

[0032] The normal tissue area is segmented according to the preset pixel group, and the pixel values ​​of each group are calculated by weighting according to the corresponding pixel weight to obtain the pixel value of each group of pixels after merging.

[0033] Arrange all merged pixels in the order of their original spatial coordinates to obtain a normal tissue image.

[0034] Preferably, the process of compressing the tumor core area and boundary area according to a scaling ratio to obtain a tumor image specifically includes:

[0035] Based on the original image, extract pixel data of the tumor core area and boundary area;

[0036] Set the tumor region scaling ratio according to the preset target low-resolution size;

[0037] Based on the scaling ratio, the tumor core area and boundary area are divided to obtain multiple feature units;

[0038] Based on the feature units, channel compression is performed using a 3D convolution kernel. After compression, the channels are arranged according to their original spatial arrangement and mapped to the corresponding coordinate positions in the low-resolution image.

[0039] Based on the mapping results, a low-resolution tumor image is obtained.

[0040] Preferably, a deep learning-based 3D reconstruction system for gynecological tumor MRI images, the system comprising:

[0041] Location module: Used to generate a 3D location box for identifying and locating minute lesions;

[0042] Segmentation module: Used to segment and output images of small lesions and images of non-small lesions containing the tumor body and normal tissue;

[0043] Boundary region generation module: used to generate the boundary region between the tumor and adjacent normal tissue;

[0044] Normal tissue generation module: used to generate normal tissue images;

[0045] Tumor generation module: used to generate low-resolution tumor images;

[0046] Image stitching module: used to stitch together lesion images, tumor images, and normal tissue images into a low-resolution image according to their original spatial coordinates;

[0047] 3D Reconstruction Module: Used to build high-precision 3D models of gynecological tumors.

[0048] Preferably, the system further includes:

[0049] The acquisition module is used to acquire raw MRI image data of clinical gynecologic oncology patients and coarse labels of tumor areas marked by doctors.

[0050] The data processing module is used to process the MRI data acquired by the acquisition module;

[0051] Storage module: Used to store the raw data acquired by the acquisition module, the processing results of the data processing module, and the intermediate data generated by each module.

[0052] The technical effects and advantages of the present invention in the above technical solution are as follows:

[0053] This invention uses a target detection model to accurately locate minute lesions and proportionally reduce their size while preserving their spatial morphology and grayscale details. It then combines grayscale difference coefficients and 3D distance weighting to generate normal tissue images, and performs 3D convolution to compress the tumor core and boundary regions, preserving structural features. Finally, image stitching yields low-resolution images adapted for model training. This allows the trained model to accurately learn the key features of minute tumor lesions, boundaries, and core regions, significantly improving the accuracy of MRI image segmentation and feature extraction. Ultimately, it constructs a high-precision 3D model of gynecological tumors, meeting the clinical needs for detailed tumor image restoration and accurate diagnosis. It is also compatible with coarse-labeled clinical data scenarios, improving processing efficiency and practical value. Attached Figure Description

[0054] Figure 1 This is a flowchart of the deep learning-based three-dimensional reconstruction method for gynecological tumor MRI images according to the present invention.

[0055] Figure 2 This is a structural diagram of the deep learning-based 3D reconstruction system for gynecological tumor MRI images according to the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] In existing technologies, multiple pixels in an MRI image are merged into one pixel through downsampling. If a small lesion on the tumor (such as a small depression or protrusion) happens to be located between these merged pixels, the small lesion is swallowed up after pixel merging. Moreover, there is a large difference in gray values ​​between the tumor boundary and normal tissue. When merging pixels of the tumor boundary and normal tissue, the gray values ​​of the boundary region are averaged, making it difficult to identify the tumor boundary. Therefore, the following method and system for three-dimensional reconstruction of gynecological tumor MRI images based on deep learning is proposed.

[0058] Example 1, please refer to... Figure 1 As shown, this embodiment provides a deep learning-based method for three-dimensional reconstruction of gynecological tumor MRI images. The steps of this deep learning-based method for three-dimensional reconstruction of gynecological tumor MRI images are explained for the specific scenario of gynecological tumor MRI image three-dimensional reconstruction:

[0059] S1: Based on the high-resolution original image, the tumor region marked with coarse labels in the original image is processed using the target detection model to generate a localization box, which is used to identify and locate small lesions.

[0060] Original image data source: The image data comes from high-resolution MRI image scans of clinical gynecological tumor patients. The MRI images are magnetic resonance images, and the scanning equipment is a 3.0T magnetic resonance imaging system used in clinical medicine.

[0061] Raw image data components: The image data includes T1-weighted image data, T2-weighted image data, and diffusion-weighted image data. These three types of image data are acquired and stored separately, including:

[0062] (1) T1-weighted image data is used to display the boundaries of anatomical structures and observe and distinguish the contours of tumors from normal tissues;

[0063] (2) T2-weighted image data, used to show areas of edema and tumor necrosis, highlighting small lesions;

[0064] (3) Diffusion-weighted image data, used to distinguish between active and necrotic areas;

[0065] The target detection model used is the improved YOLOv8-Mini, which is a model for feature extraction and anchor box modeling based on the original YOLOv8-Mini and designed for the detection of small lesions in gynecological tumor MRI images.

[0066] The reason for choosing the improved YOLOv8-Mini target detection model is that YOLOv8-Mini has strong detection capabilities for small targets. Through micro-scale anchor boxes (e.g., 2×2×2, 3×3×3 pixels), it can detect and identify tiny lesions of 1-5mm (such as tumor boundary protrusions / depressions).

[0067] The steps for generating bounding boxes using an object detection model include:

[0068] (1) Perform morphological gradient operation (e.g., Sobel operator) on the coarse label (e.g., the global binary region of the tumor marked by the doctor on the original image) to extract the approximate contour edge of the tumor and normal tissue. Using the gray value of the original image, within the 5×5×5 pixel range of the approximate contour edge, with the filtering conditions of "gray value deviates from the mean of normal tissue by ≥2 times the variance" and "spatial size ≤5mm", obtain the dataset composed of candidate areas of small lesions and generate an initial anchor box of 3×3×3 pixels.

[0069] (2) Spatial calibration, Gaussian filtering for noise reduction and gray-level normalization (i.e., uniform mapping to the range of 0-255) are performed on the original image based on the sequence of T1 weighted image data. Then, multimodal features are fused through three-dimensional convolution to generate a gray-level feature map.

[0070] (3) Use the dataset and grayscale feature map as parameters and input them into YOLOv8-Mini to train it;

[0071] (4) The trained model is used to process the tumor region marked with coarse labels on the new original image, and the model outputs the localization box corresponding to the small lesion.

[0072] S2: Based on the location results of the bounding box recognition, the original image is segmented to obtain images of small lesions and non-small lesions. The image size of the small lesion image is reduced proportionally to obtain the lesion image. The non-small lesion image includes tumors and normal tissue. In step S2, the segmentation process of the original image based on the location results of the bounding box recognition specifically includes: mapping the three-dimensional pixel coordinates of the bounding box to the original image to obtain the small lesion region in the original image; generating a mask image of the same size as the original image, and marking the small lesion region corresponding to the bounding box and the region outside the bounding box respectively; cropping the marked small lesion region in the original image based on the mask image to generate the small lesion image; and cropping the region outside the marked bounding box in the original image based on the mask image to generate the non-small lesion image.

[0073] The steps for scaling down images of small lesions include:

[0074] (1) Based on the original small lesion image obtained by segmentation, Gaussian filtering is used to remove image noise, and the three-dimensional coordinates of the positioning box are used to calibrate the edge of the lesion image to obtain the small lesion image;

[0075] (2) The scaling ratio is calculated based on the target size at low resolution and the pixel size of the small lesion;

[0076] (3) Based on the scaling ratio, the calibrated micro lesion image is input into the trilinear interpolation algorithm module. The algorithm calculates the reduced matrix based on the scaling ratio using the three-dimensional pixel matrix of the micro lesion image. The reason for choosing the trilinear interpolation algorithm is that the algorithm can calculate the gray value of each target pixel after reduction by using linear weighting based on the gray information of the surrounding pixels of the original three-dimensional pixels of the micro lesion, so that the gray gradient is smooth and the spatial morphology of the lesion is preserved.

[0077] (4) Based on the reduced matrix, back-map to the corresponding coordinate range of the three-dimensional pixel matrix of the original small lesion image, and obtain the gray values ​​of the 8 original pixels around the target pixel coordinates.

[0078] (5) By using trilinear weighting, the gray values ​​of the eight original pixels surrounding the target pixel coordinates are linearly weighted to obtain the gray value of the target pixel. After filling the gray value into the reduced matrix, the corresponding reduced micro lesion image is generated, i.e., the lesion image.

[0079] The reason for scaling down the image size of small lesion images is that tumors contain numerous small lesions with grayscale details and shapes. Scaling down (e.g., using bilinear interpolation or trilinear interpolation) can reduce the image size while maintaining the relative spatial position and grayscale distribution of the small lesion pixels.

[0080] The mask image is used for labeling. Specifically, a blank three-dimensional matrix of the same size is initialized based on the three-dimensional dimensions of the original MRI image and used as the mask image. The three-dimensional pixel coordinates of the localization box output by the target detection model are obtained. Pixel values ​​within the localization box in the mask image are marked as 1, indicating small lesion areas. Pixel values ​​outside the localization box in the mask image are marked as 0, indicating areas other than small lesion areas, including tumors and healthy normal tissue.

[0081] To verify the accuracy of the small lesion areas marked by the mask, the areas with a pixel value of 1 in the mask image were extracted, a three-dimensional coordinate range was generated, and the range was compared with the positioning box.

[0082] S3: Based on the original image, obtain the image grayscale values, fuse them to generate a grayscale feature map, generate boundary candidate points according to the grayscale gradient of the grayscale feature map, cluster the boundary candidate points in three-dimensional space to generate boundary regions, calculate the minimum bounding cube of each boundary region to obtain the boundary area, where the boundary area includes the tumor boundary and the adjacent normal tissue boundary. In step S3, the process of generating boundary candidate points according to the grayscale gradient of the grayscale feature map specifically includes:

[0083] Based on the grayscale feature map, calculate the three-dimensional grayscale gradient of each pixel, and use the pixels with non-zero gradient values ​​to generate initial boundary candidate points.

[0084] Based on the gray-scale difference coefficient, a dynamic threshold is set, and the initial boundary candidate points are filtered by comparing the gradient value with the dynamic threshold.

[0085] Based on the filtered candidate points, neighborhood analysis is used to obtain consecutive pixels in the neighborhood that have at least two candidate points of the same type.

[0086] Based on the tumor region marked by the coarse label, the boundary between the tumor region and normal tissue is used as a limiting condition to filter continuous pixels in the neighborhood to obtain boundary candidate points.

[0087] Gray-scale feature map generation: Gray-scale values ​​are obtained from the original image data. These gray-scale values ​​are then fused to obtain a gray-scale feature map. Specifically, from clinically acquired MRI images, including T1-weighted, T2-weighted, and diffusion-weighted images stored in DICOM format, the original gray-scale values ​​of each image are extracted and converted into a three-dimensional pixel matrix. Each matrix element corresponds to the gray-scale value of a single pixel. The gray-scale value matrices of each image are preprocessed. A multimodal gray-scale fusion module is constructed, using three sets of 1×1×1 three-dimensional convolutional kernels to compress the channels of the preprocessed gray-scale matrices of the three image data. Based on the tumor feature representation ability of each image data, the compressed feature vectors are weighted and summed. The weighted fused feature vectors are then reconstructed into a three-dimensional matrix, thereby generating the gray-scale feature map.

[0088] The grayscale matrix of each image data is preprocessed, specifically by uniformly mapping the grayscale values ​​of the image data (the mapping range is 0-255), and then smoothing the mapped grayscale matrix by using a Gaussian filter with standard deviation (σ=1.0).

[0089] The gradient value is compared with the dynamic threshold. The specific comparison process includes: based on the acquired initial boundary candidate points and the dynamic threshold, traversing all initial boundary candidate points, extracting the three-dimensional gray-level gradient magnitude of each candidate point, and then comparing the gradient magnitude of each candidate point with the dynamic threshold. Candidate points with gradient magnitude greater than or equal to the dynamic threshold are retained (higher gradient magnitude means more significant gray-level changes in the candidate point), and candidate points with gradient magnitude less than the dynamic threshold are removed. The retained candidate points are verified, and those within the effective pixel range of the original image are selected. After selection, a set of boundary candidate points is obtained.

[0090] The dynamic threshold is obtained by multiplying the gray-level difference coefficient between the tumor and normal tissue with the gradient mean of the gray-level feature map. The larger the K value and the higher the gradient mean, the higher the dynamic threshold.

[0091] Neighborhood analysis yields continuous pixels. Specifically, for the set of boundary candidate points after threshold filtering, a 3×3×3 three-dimensional neighborhood is used as a benchmark. Boundary candidate points located within the 3×3 neighborhood are selected from the set of boundary candidate points. For each boundary candidate point, the coordinates of all pixels within its 3×3×3 neighborhood are extracted, and matched with the filtered set of boundary candidate points. The number of pixels belonging to the same type of candidate point (pixels in the same filtered set) within the same neighborhood is analyzed and counted. Based on the filtering criteria, continuous pixels that meet the criteria are obtained. Continuous pixels indicate that they are spatially continuous with other candidate points.

[0092] Among them, the screening conditions include: (1) If the number of similar candidate points in the neighborhood of a candidate point is ≥2, then the candidate point is determined to be a continuous pixel; if the number of similar candidate points in the neighborhood of a candidate point is <2, then it is determined to be a non-continuous pixel.

[0093] In step S3, the minimum circumscribed cuboid of each boundary region is calculated. The specific process of obtaining the boundary region includes:

[0094] Obtain the 3D coordinates of all pixels within each boundary region after clustering in 3D space;

[0095] Calculate the maximum and minimum values ​​on the x, y, and z axes based on the three-dimensional coordinates of each pixel in the boundary region;

[0096] The starting vertex is generated by finding the minimum values ​​of the x, y, and z axes of each boundary region, and the diagonal vertex is generated by finding the maximum values ​​of the x, y, and z axes of each boundary region. Based on the starting vertex and the diagonal vertex, the minimum circumscribed cuboid is constructed.

[0097] The boundary region is obtained by determining the spatial extent of the smallest circumscribed cuboid corresponding to each boundary region;

[0098] Boundary candidate points are clustered. By analyzing the neighborhood of the boundary candidate points, they are converted into a 3D point cloud dataset corresponding to the spatial location of the grayscale feature map. The DBSCAN algorithm is used (with a neighborhood radius of ε = 2-3 pixels and a minimum number of core points MinPts = 3-5). The core points are traversed, the point cloud is marked, and the clusters are recursively clustered. Finally, small clusters with a spatial distance < ε are merged, and invalid clusters with a number of points < MinPts are removed to obtain a spatially continuous boundary region.

[0099] S4: Select samples from the core area of ​​the tumor and samples from the normal tissue far from the boundary area, calculate the gray-level mean for each, and calculate the gray-level difference coefficient between the tumor and normal tissue based on the gray-level difference coefficient. Then, weight the normal tissue based on the distance to the boundary area and merge the pixels to obtain the normal tissue image. Specifically, step S4 involves the following steps:

[0100] Calculate the three-dimensional spatial distance between each pixel and the nearest boundary region for each pixel within the normal tissue region;

[0101] Pixel weights are assigned to normal tissue regions based on grayscale difference coefficients and three-dimensional spatial distance.

[0102] The normal tissue area is segmented according to the preset pixel group, and the pixel values ​​of each group are calculated by weighting according to the corresponding pixel weight to obtain the pixel value of each group of pixels after merging.

[0103] Arrange all the merged pixels in the order of their original spatial coordinates to obtain a normal tissue image;

[0104] Each pixel group is weighted and the pixels are merged to obtain the pixel value. Specifically, for each preset pixel group, the original grayscale value and corresponding pixel weight of all pixels in the pixel group are extracted. Each pixel in each pixel group is weighted and then the sum of the weights of all pixels in the group is calculated to obtain the pixel value of the merged pixel group.

[0105] S5: Compress the tumor core area and boundary area according to the scaling ratio to obtain the tumor image. Step S5 specifically includes the following steps:

[0106] Based on the original image, extract pixel data of the tumor core area and boundary area;

[0107] Set the tumor region scaling ratio according to the preset target low-resolution size;

[0108] Based on the scaling ratio, the tumor core area and boundary area are divided to obtain multiple feature units;

[0109] Based on the feature units, channel compression is performed using a 3D convolution kernel. After compression, the channels are arranged according to their original spatial arrangement and mapped to the corresponding coordinate positions in the low-resolution image.

[0110] Based on the mapping results, a low-resolution tumor image is obtained;

[0111] The criteria for defining the tumor core area are: (1) a tumor area that is ≥3-5 pixels away from the boundary area; (2) a pixel area whose gray value is within the range of "mean gray value of tumor core area sample ±5%" based on criterion (1).

[0112] S6: The lesion image, tumor image, and normal tissue image are stitched together according to their original size ratio and spatial coordinates to obtain a low-resolution image. The low-resolution image is used to train the model. Based on the trained model, the MRI image is segmented and features are extracted to construct a three-dimensional model of the tumor.

[0113] Example 2: To implement the method of Example 1, please refer to... Figure 2 As shown, this embodiment provides a deep learning-based three-dimensional reconstruction system for gynecological tumor MRI images. The system includes:

[0114] The acquisition module is used to acquire raw MRI image data of clinical gynecologic oncology patients and coarse labels of tumor areas marked by doctors.

[0115] The data processing module is used to process the MRI data acquired by the acquisition module;

[0116] Storage module: Used to store the raw data acquired by the acquisition module, the processing results of the data processing module, and the intermediate data generated by each module;

[0117] The localization module is used to generate three-dimensional bounding boxes for identifying and locating small lesions. It processes the tumor regions marked with coarse labels in the original image through a target detection model to generate bounding boxes.

[0118] Segmentation module: Used to segment and output images of small lesions and images of non-small lesions containing the tumor body and normal tissue. It identifies the localization results through the positioning box, segments the original image, and obtains images of small lesions and non-small lesions.

[0119] Boundary region generation module: used to generate the boundary region between the tumor and adjacent normal tissue. It obtains the gray value of the image based on the original image, and generates a gray-scale feature map after fusion. Based on the gray-scale gradient of the gray-scale feature map, it generates boundary candidate points, clusters the boundary candidate points in three-dimensional space to generate boundary regions, and calculates the minimum bounding cube of each boundary region to obtain the boundary region.

[0120] Normal tissue generation module: used to generate normal tissue images. It selects samples from the core area of ​​the tumor and samples from the normal tissue far from the boundary area, calculates the gray-level mean, calculates the gray-level difference coefficient between the tumor and normal tissue based on the gray-level mean, and weights the normal tissue according to the distance between the normal tissue and the boundary area based on the gray-level difference coefficient. After pixel merging, the normal tissue image is obtained.

[0121] Tumor generation module: Used to generate low-resolution tumor images by compressing the tumor core area and boundary area according to a scaling ratio to obtain the tumor image;

[0122] Image stitching module: used to stitch lesion images, tumor images and normal tissue images into a low-resolution image according to the original spatial coordinates. This is done by stitching lesion images, tumor images and normal tissue images according to the original size ratio and spatial coordinates.

[0123] 3D Reconstruction Module: Used to construct high-precision 3D models of gynecological tumors. It trains the model using low-resolution images, and then segments and extracts features from MRI images based on the trained model to construct a 3D model of the tumor.

[0124] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for three-dimensional reconstruction of gynecological tumor MRI images based on deep learning, characterized in that, The method includes: Based on the high-resolution original image, the tumor region marked with coarse labels in the original image is processed using a target detection model to generate a localization box, wherein the localization box is used to identify and locate small lesions. Based on the localization results of the bounding box recognition, the original image is segmented to obtain images of small lesions and non-small lesions. The image size of the small lesion image is reduced proportionally to obtain the lesion image. The non-small lesion image includes tumors and normal tissue. Based on the original image, the grayscale values ​​of the image are obtained, and a grayscale feature map is generated after fusion. Based on the grayscale gradient of the grayscale feature map, boundary candidate points are generated. The boundary candidate points are clustered in three-dimensional space to generate boundary regions. The minimum bounding cube of each boundary region is calculated to obtain the boundary region, wherein the boundary region includes the tumor boundary and the adjacent normal tissue boundary. Samples from the core area of ​​the tumor and samples from normal tissue far from the boundary area were selected, and the mean gray value was calculated for each. Based on the mean gray value, the gray value difference coefficient between the tumor and normal tissue was calculated. Based on the gray value difference coefficient and weighted according to the distance between the normal tissue and the boundary area, the normal tissue image was obtained after pixel merging. The tumor core and boundary regions are compressed according to the scaling ratio to obtain the tumor image; The lesion image, tumor image, and normal tissue image are stitched together according to their original size ratio and spatial coordinates to obtain a low-resolution image. The low-resolution image is used to train a model, and the trained model is used to segment and extract features from the MRI image to construct a three-dimensional model of the tumor.

2. The method for three-dimensional reconstruction of gynecological tumor MRI images based on deep learning according to claim 1, characterized in that, The process of segmenting the original image based on the localization result of the bounding box recognition specifically includes: The three-dimensional pixel coordinates of the positioning box are mapped to the original image to obtain the small lesion area in the original image; Generate a mask image of the same size as the original image, and mark the small lesion area corresponding to the positioning box and the area outside the positioning box respectively; Based on the mask image, the marked small lesion regions in the original image are cropped to generate small lesion images; Based on the mask image, the region outside the marked location box in the original image is cropped to generate an image of non-micro lesions.

3. The method for three-dimensional reconstruction of gynecological tumor MRI images based on deep learning according to claim 1, characterized in that, The process of generating boundary candidate points based on the gray-level gradient of the gray-level feature map specifically includes: Based on the grayscale feature map, calculate the three-dimensional grayscale gradient of each pixel, and use the pixels with non-zero gradient values ​​to generate initial boundary candidate points. Based on the gray-scale difference coefficient, a dynamic threshold is set, and the initial boundary candidate points are filtered by comparing the gradient value with the dynamic threshold. Based on the filtered candidate points, neighborhood analysis is used to obtain consecutive pixels in the neighborhood that have at least two candidate points of the same type. Based on the tumor region marked by the coarse label, the boundary between the tumor region and normal tissue is used as a limiting condition to filter continuous pixels in the neighborhood to obtain boundary candidate points.

4. The method for three-dimensional reconstruction of gynecological tumor MRI images based on deep learning according to claim 1, characterized in that, The process of calculating the minimum circumscribed cuboid of each boundary region to obtain the boundary region specifically includes: Obtain the 3D coordinates of all pixels within each boundary region after clustering in 3D space; Calculate the maximum and minimum values ​​on the x, y, and z axes based on the three-dimensional coordinates of each pixel in the boundary region; The starting vertex is generated by finding the minimum values ​​of the x, y, and z axes of each boundary region, and the diagonal vertex is generated by finding the maximum values ​​of the x, y, and z axes of each boundary region. Based on the starting vertex and the diagonal vertex, the minimum circumscribed cuboid is constructed. The boundary region is obtained by determining the spatial extent of the smallest circumscribed cuboid corresponding to each boundary region.

5. The method for three-dimensional reconstruction of gynecological tumor MRI images based on deep learning according to claim 1, characterized in that, The process of obtaining a normal tissue image by merging pixels based on a gray-level difference coefficient and weighting according to the distance between normal tissue and the boundary region specifically includes: Calculate the three-dimensional spatial distance between each pixel and the nearest boundary region for each pixel within the normal tissue region; Pixel weights are assigned to normal tissue regions based on grayscale difference coefficients and three-dimensional spatial distance. The normal tissue area is segmented according to the preset pixel group, and the pixel values ​​of each group are calculated by weighting according to the corresponding pixel weight to obtain the pixel value of each group of pixels after merging. Arrange all merged pixels in the order of their original spatial coordinates to obtain a normal tissue image.

6. The method for three-dimensional reconstruction of gynecological tumor MRI images based on deep learning according to claim 1, characterized in that, The process of compressing the tumor core and boundary regions according to a scaling ratio to obtain a tumor image specifically includes: Based on the original image, extract pixel data of the tumor core area and boundary area; Set the tumor region scaling ratio according to the preset target low-resolution size; Based on the scaling ratio, the tumor core area and boundary area are divided to obtain multiple feature units; Based on the feature units, channel compression is performed using a 3D convolution kernel. After compression, the channels are arranged according to their original spatial arrangement and mapped to the corresponding coordinate positions in the low-resolution image. Based on the mapping results, a low-resolution tumor image is obtained.

7. A deep learning-based three-dimensional reconstruction system for gynecological tumor MRI images, characterized in that, The system includes: The localization module is used to generate three-dimensional bounding boxes for identifying and locating small lesions. It processes the tumor regions marked with coarse labels in the original image through a target detection model to generate bounding boxes. Segmentation module: Used to segment and output images of small lesions and images of non-small lesions containing the tumor body and normal tissue. It identifies the localization results through the positioning box, segments the original image, and obtains images of small lesions and non-small lesions. Boundary region generation module: used to generate the boundary region between the tumor and adjacent normal tissue. It obtains the gray value of the image based on the original image, and generates a gray-scale feature map after fusion. Based on the gray-scale gradient of the gray-scale feature map, it generates boundary candidate points, clusters the boundary candidate points in three-dimensional space to generate boundary regions, and calculates the minimum bounding cube of each boundary region to obtain the boundary region. Normal tissue generation module: used to generate normal tissue images. It selects samples from the core area of ​​the tumor and samples from the normal tissue far from the boundary area, calculates the gray-level mean, calculates the gray-level difference coefficient between the tumor and normal tissue based on the gray-level mean, and weights the normal tissue according to the distance between the normal tissue and the boundary area based on the gray-level difference coefficient. After pixel merging, the normal tissue image is obtained. Tumor generation module: Used to generate low-resolution tumor images by compressing the tumor core area and boundary area according to a scaling ratio to obtain the tumor image; Image stitching module: used to stitch lesion images, tumor images and normal tissue images into a low-resolution image according to the original spatial coordinates. This is done by stitching lesion images, tumor images and normal tissue images according to the original size ratio and spatial coordinates. 3D Reconstruction Module: Used to construct high-precision 3D models of gynecological tumors. It trains the model using low-resolution images, and then segments and extracts features from MRI images based on the trained model to construct a 3D model of the tumor.

8. The deep learning-based three-dimensional reconstruction system for gynecological tumor MRI images according to claim 7, characterized in that, The system also includes: The acquisition module is used to acquire raw MRI image data of clinical gynecologic oncology patients and coarse labels of tumor areas marked by doctors. The data processing module is used to process the MRI data acquired by the acquisition module; Storage module: Used to store the raw data acquired by the acquisition module, the processing results of the data processing module, and the intermediate data generated by each module.

Images