Prostate contouring method based on geometric intersection over union loss

Abstract

The application discloses a prostate contour segmentation method based on a geometric intersection-over-union loss, and comprises the following steps: constructing a geometric intersection-over-union loss function based on a Gaussian area formula, combining a traditional distance loss to construct an organ edge contour regression model, adopting the contour regression model to model an organ segmentation task, and modeling the organ segmentation task in a mode of reconstructing an organ contour layer by layer by using an axial surface slice organ contour. The contour regression model is composed of a graph convolution network and a multilayer perception mechanism to form a multilayer contour regression neural network. The application can model a segmentation task in an edge contour regression reconstruction mode, the modeling mode is simple and effective, the result is accurate, the false positive prediction of a classification probability prediction model based on a UNet or the like encoder-decoder structure can be reduced, a segmentation result with better edge consistency and better continuity is obtained, the influence of similar textures or fuzzy boundaries on a segmentation model is reduced, and the accuracy of the regression model is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of medical image processing and application, and particularly relates to a prostate contour segmentation method based on a geometric intersection-over-union loss. BACKGROUND

[0002] As a key link of a computer-aided detection (CADe) system, an accurate and effective automatic prostate segmentation algorithm can accurately outline the organ shape in MR images. It plays an important role in assisting in diagnosis and decision-making, preoperative planning, puncture guidance, radiotherapy planning and early detection screening, and has important research significance. In recent years, artificial neural networks have made breakthrough progress in the field of computer vision and have been widely used in various visual downstream tasks. Subsequently, they have been gradually applied to the segmentation task of medical images. The characteristics and advantages are to use the original image data, use a convolutional neural network to automatically extract features, and make binary classification probability prediction for each voxel in a three-dimensional body data field to realize end-to-end automatic segmentation. The convolutional neural network based on an encoder-decoder structure has been widely used in prostate segmentation of MR images and has become the mainstream algorithm in this field.

[0003] However, the pixel classification model is a computationally intensive model that needs to calculate the multi-level feature map of each point. Limited by the GPU memory, this kind of method often needs to divide the data into regional patches and use a sliding window to process the entire data. The cutting and cropping will affect the accuracy of positioning and the fineness of segmentation, and lack complete global information. Limited by the receptive field, the convolutional network is difficult to use global information and establish long-range dependency, and the pixel classification task is also easily affected by the imbalance of labeled categories and similar textures and gray patterns, resulting in discrete boundaries and poor continuity. The improvement of the network output depends on the post-processing method. A network model with strong combination of global information and use of local information is needed to improve the current full convolution model prediction result consistency and the problem of easy occurrence of discrete false positive prediction. SUMMARY

[0004] The purpose of the present application is to provide a prostate contour segmentation method based on a geometric intersection-over-union loss to solve the problems existing in the prior art.

[0005] To achieve the above purpose, the present application provides a prostate contour segmentation method based on a geometric intersection-over-union loss, comprising:

[0006] Collecting a T2-weighted MR image, preprocessing the T2-weighted MR image into an initial axial slice data set;

[0007] Judging whether the axial slice in the initial axial slice data set contains a prostate region, and obtaining a regression slice data set containing the prostate region;

[0008] detecting and positioning the prostate region in the regression slice dataset to obtain the prostate center point coordinate and the width and height of the detection box;

[0009] constructing a contour regression network model and training the model;

[0010] obtaining an initial contour point set based on the prostate center point coordinate and the width and height of the detection box, and obtaining a predicted contour point set through the contour regression network model;

[0011] reconstructing a binary mask layer by layer based on the predicted contour point set to obtain a prostate organ three-dimensional segmentation result.

[0012] Optionally, the T2-weighted MR image preprocessing process comprises: correcting the T2-weighted MR image using an N4 bias field, performing adaptive histogram equalization, and then performing normalization operation; normalizing the MR image, slicing the normalized MR image to obtain a continuous axial surface sequence image; and performing offline data enhancement and dynamic data enhancement on the continuous axial surface sequence image to obtain an initial axial surface slice dataset, wherein the slices in the dataset are two-dimensional slices.

[0013] Optionally, the regression slice dataset obtaining process comprises: identifying whether the axial surface slice in the initial axial surface slice dataset contains a prostate region through a classification network, the image input into the classification network being three adjacent slices in the axial surface, if the prostate region is contained, the slice is determined as a regression slice, and if the prostate region is not contained, the slice is removed without subsequent operation.

[0014] Optionally, the detecting and positioning process comprises: using a CenterNET architecture taking a DLA network as a backbone extraction network, using three consecutive slices before and after in the regression slice dataset as input images to perform detecting and positioning, and obtaining the prostate target detection center point coordinate and the width and height of the detection positioning box.

[0015] Optionally, the initial contour point set construction process comprises: obtaining the midpoint coordinates of four boundary edges of the detection box through the prostate target detection center point coordinate and the width and height data of the detection positioning box, correcting four edge points of the prostate region based on the boundary midpoint coordinates, constructing an octagonal contour in a clockwise order with the corrected edge points as the center, representing the octagonal contour in a directed graph form, and constructing the initial contour point set.

[0016] Optionally, the contour regression network model is composed of a shared weight Shared-MLP and a feature sequence module, the feature sequence module is composed of 1 CMixer module without residual connection and 7 consecutive residual Res-CMixer modules, the CMixer module is composed of a neighborhood feature aggregation module and a channel mixing submodule, the neighborhood aggregation submodule is composed of a circular convolution module, and the channel mixing module is composed of a Token Mixer structure module.

[0017] Optionally, the process of training the model comprises: constructing a geometric reconstruction loss function, constraining the model training process based on the geometric reconstruction loss function, obtaining the area surrounded by the predicted polygon contour and the gold standard contour, and maximizing the intersection area of the area surrounded by the predicted polygon contour and the gold standard contour as the optimization target to constrain the model to update the weight, wherein the geometric reconstruction loss function is composed of a loss based on geometric intersection over union and a loss based on distance.

[0018] Optionally, the process of obtaining the predicted contour point set comprises: inputting the feature sequence of the initial contour position corresponding to the feature set extracted by the backbone network after the regression slice data set into the contour regression network model, obtaining the coordinate offset of each point of the initial contour, and adding the coordinate offset to the initial contour coordinate to obtain the predicted contour point set.

[0019] Optionally, the process of obtaining the prostate organ three-dimensional segmentation result comprises: constructing the predicted contour point set into a boundary contour, filling the contour inside by using computer vision technology, reconstructing a binary segmentation mask layer by layer, and stacking to the original position to reconstruct a three-dimensional segmentation result.

[0020] The technical effects of the present application are:

[0021] (1) The segmentation network is constructed based on a regression model, which can avoid the discrete and broken prediction results of the pixel-by-pixel classification of the convolution network, and the edge continuity and consistency are better;

[0022] (2) The basic regression network module based on circular convolution and MLP-Mixer can combine the neighborhood feature capture ability and the global information modeling ability, and compared with using only circular convolution or MLP-Mixer structure, better precision performance is obtained;

[0023] (3) The shape perception loss function based on the Gaussian area formula can utilize the overall shape information of the contour coordinates to encode the target, so that the regression model is more accurate and efficient;

[0024] (4) The segmentation model provided by this method has high accuracy and practicality. The method is simple to operate. The automated segmentation method can save the time of manual delineation in clinical practice. It has high accuracy and practicality, reduces delineation costs, assists in clinical diagnosis, and has practical value. Attached Figure Description

[0025] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0026] Figure 1 This is a flowchart illustrating the overall framework of an embodiment of the present invention.

[0027] Figure 2 This is a schematic diagram of the network structure in an embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram of the basic network module in an embodiment of the present invention;

[0029] Figure 4 This is a schematic diagram illustrating the calculation of the area of ​​any polygon involved in the loss function in an embodiment of the present invention;

[0030] Figure 5 This is a visual schematic diagram of the prostate segmentation results in an embodiment of the present invention;

[0031] Figure 6 This is a schematic diagram of the prostate segmentation result outline in an embodiment of the present invention;

[0032] Figure 7 This is a schematic diagram showing the comparison between the ProstateX data in this embodiment of the invention, concentrated in the slice range containing the prostate region, and the current convolution method. Detailed Implementation

[0033] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0034] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0035] Example 1

[0036] like Figures 1-6 As shown, this embodiment provides a prostate contour segmentation method based on geometric intersection-union ratio loss, including:

[0037] The embodiment relates to a method for prostate organ segmentation in T2-weighted MR images. According to the embodiment, a prostate segmentation algorithm based on a geometric intersection-over-union loss can accurately delineate a prostate organ region in a T2-weighted MR image, constrain a regression network through a geometric reconstruction loss, realize accurate prostate region segmentation, and assist in diagnosis, detection and identification of prostate-related diseases.

[0038] Figure 1 The prostate segmentation algorithm based on the geometric intersection-over-union loss provided by the embodiment is specifically shown in the following steps:

[0039] Step 1, a classification network such as DenseNet, Vision Transformer or the like is used to perform image identification on a preprocessed prostate image, and a slice S containing a prostate is screened out Valid for ROI positioning and image feature extraction. The image preprocessing step includes N4 bias field correction of the image, adaptive image histogram equalization, resampling of the image to a fixed scale, processing into continuous axial surface two-dimensional slices and offline data augmentation, including horizontal mirror flipping, adaptive Gaussian noise interference, motion blur, affine transformation, optical distortion, grid distortion, elastic transformation, adaptive histogram equalization, image sharpening, random brightness and contrast enhancement and online data augmentation of affine transformation. The image input into the identification network is three adjacent slices in the axial surface, so as to inherit the pre-training weight of the classification model in the natural image dataset, and the axial surface slice determined as not containing the target region does not participate in the subsequent regression task;

[0040] Step 2, a positioning network is used to detect and position the ROI region. The prostate-containing slice screened out in step 1 is input into a target detection network Φ(I(x); θ) with the three consecutive slices before and after the slice as input images I(x), wherein the detection network adopts a CenterNet architecture and uses DLA34 or a more complex DLA60 as a backbone network responsible for feature extraction for the target detection network and the subsequent regression network. The network predicts and outputs the center point coordinates c of the prostate in the slice, and the width w and the height h of the detection positioning box, and simultaneously extracts image features χ from the backbone network. 2

[0041] ​Step 3, the midpoint of the target positioning package frame four edges is surrounded, and the regression network is used to correct the coordinates of the four edge points of the leftmost end l, the topmost end t, the rightmost end r, and the bottommost end b of the ROI, representing the four most boundary edge points of the prostate region, called extreme points. In a clockwise order, an octagonal contour is constructed by linear interpolation based on the edge points and the initial contour C0 is constructed in the form of a directed graph.

[0042] Step 4, feature sampling is performed according to the initial contour C0, and the feature vector corresponding to the position of the backbone network multi-level feature χ is extracted according to the contour coordinate position, and is converted into an ordered feature vector sequence V.

[0043] Step 5, contour point set regression is performed using the contour regression network: that is, the feature sequence V and the coordinate position obtained by sampling in step 4 are used to perform point-to-point boundary contour point set regression on the initial contour, and the coordinate offset C of the predicted contour relative to the initial contour point set is obtained. offset .

[0044] As shown in Figure 2 , the regression network is a PointNet-like network model composed of basic modules composed of multiple layers of shared MLP, ring convolution, and MLP-Mixer. The CMixer sequence is composed of 1 CMixer module without residual connection and 7 consecutive residual CMixer modules, which are used for feature mapping, neighborhood aggregation, and capturing global information of the feature sequence. Finally, the contour coordinate regression is performed by the residual CMixer module and the sequence shared weight MLP. The CMixer module is composed of a neighborhood aggregation constructed by ring convolution and a channel mixing based on the MLP-Mixer structure, which is used for local information aggregation and long-distance dependence relationship establishment. The neighborhood aggregation submodule is composed of a ring graph convolution GCN module, which is used for local neighborhood feature aggregation of adjacent control point features, and is defined as follows:

[0045] Y i =Norm(σ(V i ·W+b))i=1,2,…,N

[0046] Where V j is the neighborhood feature vector at position j, W is the learnable weight of the corresponding position and channel, and N is the number of contour control points. The channel mixing module is composed of an MLP-Mixer module, which is used to construct the information propagation path between all nodes, establish long-distance dependence relationship, and improve the global perception ability of the regression network model. The related calculation is defined as follows:

[0047] Y=T -1 (Norm(σ(T(X)·W+b)))

[0048] where T is the transpose operation, σ(·) is the GELU function, Norm(·) is the layer normalization operation. The regression network is constrained by a geometry reconstruction loss function in the training stage. In the geometry reconstruction loss, the loss function based on the geometric intersection-over-union is used as the shape-aware loss function, and the triangular form based on the Gaussian area formula is used to calculate the area of any polygon, as shown in the following formula: Figure 3 The overall polygon can be decomposed into a series of equilateral triangles composed of adjacent contour nodes and origin points, such as △OAB, and the area calculation is expressed as:

[0049]

[0050] and a series of negative triangles opposite to the positive direction, such as △OCD, and the area can be expressed as:

[0051]

[0052] The overall area A(P i ) of the polygon is the sum of the areas of all positive triangles and negative triangles, expressed as:

[0053]

[0054] The geometric intersection-over-union loss function is constructed in the form of negative log-likelihood term:

[0055]

[0056] where P represents the predicted polygon contour point set, i.e., the predicted contour, and the predicted polygon contour is obtained based on the predicted contour point set. P gt is the ground truth polygon contour point set obtained by manual labeling by doctors in the training process, P A ∩P B represents the intersection polygon contour point set extraction, and Shapely is used to extract the intersection polygon contour point set between the predicted polygon and the ground truth polygon. The distance regression loss in the geometry reconstruction loss adopts distance loss, which is used for positioning and clockwise arrangement order constraint:

[0057]

[0058] The geometry reconstruction loss of the regression network is finally the weighted sum of the shape-aware loss and the distance loss function, where λ = 1 is the balance coefficient, which is used to make the losses of the two have similar overall penalties.

[0059] L Polygon =L GeoIoU +λL smooth L1

[0060] Step 6, reconstruction of regression results:

[0061] For the output results of the regression network, the predicted contour point set after regression is used as the boundary contour, and the boundary control points are filled inside the contour using the computer vision library OpenCV to reconstruct the binary segmentation mask. Through verification in the ProstateX dataset, compared with the traditional convolution method, the segmentation DSC index of the current method is improved by 0.89-3.93% and 0.84-3.88% under two kinds of backbone networks. The geometric reconstruction loss is constructed by using the loss function based on the geometric intersection over union combined with the distance loss, which is improved by 1.37-3.23% and 0.16-1.61% compared with using only the distance loss or the geometric intersection over union loss based on the DSC index. The comparison results of the segmentation index of the current method and the convolution-based method in the range of the prostate slice are shown in Table 1. The comparison results of the regression network constrained by different loss functions are shown in Table 2.

[0062] Table 1

[0063]

[0064] Table 2

[0065]

[0066] Embodiment Two

[0067] In this embodiment, a prostate contour segmentation method based on geometric intersection over union loss is provided, which comprises:

[0068] (1) Image preprocessing: that is, the T2 weighted MR image collected in clinic is preprocessed, including resampling to a uniform size, performing N4 bias field correction and adaptive histogram equalization operation, normalizing the image, and then cutting the MR image into a continuous axial slice sequence image, performing offline data enhancement including horizontal mirror flip, adaptive Gaussian noise interference, motion blur, affine transformation, optical distortion, grid distortion, elastic transformation, adaptive histogram equalization, image sharpening, random brightness and contrast enhancement, and online data enhancement operation including affine transformation, to expand the data set and avoid overfitting of the network during training;

[0069] (2) A general classification network is used to identify whether each axial slice contains a prostate region. The axial slice that is judged to not contain the target region does not participate in the subsequent regression task;

[0070] (3) A positioning network is used to coarsely position the prostate region in the axial slice: that is, CenterNet with DLA network as the backbone network is used to coarsely position the prostate region to obtain the target center point coordinates and the width and height of the bounding box;

[0071] (4) Initial polygon contour construction and feature extraction: The midpoints of the four edges of the target positioning bounding box are used to correct the leftmost, uppermost, rightmost, and lowermost edge points of the prostate target. The eight initial contour polygons of the edge regression are constructed in a clockwise order with the edge points as the centers in a directed graph form. The feature vectors of the corresponding positions of the backbone network multi-level features are extracted according to the contour control point coordinate positions;

[0072] (5) Contour point set regression using a contour regression network: The feature sequence and coordinate positions obtained by sampling in step (3) are used to perform point-to-point boundary contour point set regression on the initial contour to obtain the coordinate offset of the predicted contour point set relative to the initial contour point set;

[0073] The regression network, as shown in the figure, is a PointNet-like artificial neural network model composed of basic modules composed of multi-layer shared weight MLP, ring convolution, and MLP-Mixer combination; Figure 2

[0074] (6) In the training process of the network in step (5) above, the geometric intersection-over-union-based loss function and the distance-based loss function are used together to form a geometric reconstruction loss function to constrain the model training process. The intersection area between the predicted contour of the contour regression network and the gold standard contour is maximized as the optimization objective to constrain the network weight update.

[0075] (7) Input the test image into the regression network to obtain the coordinate offset of each point in the target contour point set;

[0076] (8) Add the initial contour point set coordinates to the corresponding coordinate offset of each point to obtain the coordinates of the regression contour point set;

[0077] (9) Use the regression contour point set as the target predicted contour to reconstruct the binary mask layer by layer and stack to generate a three-dimensional segmentation result.

[0078] Example Three

[0079] The prostate contour segmentation method based on geometric intersection-over-union loss provided in this embodiment includes:

[0080] ​The prostate segmentation method based on regression reconstruction proposed in the embodiment adopts a polygon reconstruction loss function constituted based on a geometric IoU loss function as the contour regression model of the method, and the design principle of the loss function is: the overall area of the intersection polygon is calculated by using the regression polygon contour coordinates and the gold standard contour coordinates, and the prediction result with a smaller IoU is punished more by using the negative logarithmic constraint form, the shape information is encoded by using the contour coordinates, the shape perception ability of the network is improved, so that the regression network has better shape fitting performance, and the overall positioning and smoothing of the network prediction contour boundary are performed by using the distance loss, so as to reduce the prediction deviation of the network model, improve the segmentation accuracy of the overall model, and make the model prediction edge contour accurately fit the target boundary contour. Experimental results show that the algorithm is efficient and accurate.

[0081] To achieve the above purpose, specifically, the prostate organ segmentation method based on boundary contour regression reconstruction of the embodiment includes the following steps: image preprocessing based on T2 weighted MR image, and processing into 2D slices. Use a classification network to identify effective slices. Use a target detection network to locate the target region (Region of Interest, ROI) and extract features. Generate an initial contour and extract features. Use a contour regression network to regress point-to-point contour boundary point sets, and predict the coordinate offset between the initial contour and the predicted contour. In the training process of the network, a geometric IoU loss function is constructed in the form of a triangular Gaussian area formula, and a SmoothL1 loss function is used to constrain the contour regression. Based on the regression result, a binary mask is reconstructed, and a three-dimensional segmentation entity is stacked.

[0082] As above, wherein the image preprocessing step includes using N4 bias field correction on the MR image, performing gray histogram equalization, resampling image scaling, and scaling the MR to a fixed scale, processing into continuous axial two-dimensional slices, and performing offline data enhancement, including horizontal mirror flip, adaptive Gaussian noise interference, motion blur, affine transformation, optical distortion, grid distortion, elastic transformation, adaptive histogram equalization, image sharpening, random brightness and contrast enhancement, and online data enhancement of affine transformation.

[0083] As above, wherein the effective slice identification step is to use a general image classification network such as ResNet and VGG network to identify and screen the slices containing prostate, and the screened slices are used for subsequent feature extraction.

[0084] As above, wherein the target detection network uses the CenterNet framework to identify the RoI region. Meanwhile, the DLA backbone network in the detection network is used to extract image features, and the feature sampling is performed at the corresponding coordinate position by using the initial contour point set coordinates.

[0085] The profile regression network adopts a continuous shared weight MLP and a CMixer module composed of a ring convolution and an MLP-Mixer stack, and performs profile coordinate position offset vector regression on the corresponding feature vectors of the input profile point coordinates, and the predicted control point coordinate offset is used to calculate the predicted profile coordinates formed from the initial profile.

[0086] The profile regression loss function of the constraint network regression is composed of a distance loss and a shape-aware loss based on a geometric intersection-over-union. The distance regression loss is used for profile coordinate positioning and edge smoothing, and a SmoothL1 loss function is adopted. The shape-aware loss is used to encode the overall geometric shape information using the profile control point coordinates, maximize the proportion of the intersection polygon, and adopt a triangular form of the Gaussian area formula and construct in the form of a negative logarithm, so that smaller intersection areas are punished with greater loss and obtain a greater update gradient.

[0087] The binary segmentation mask is reconstructed using the predicted profile coordinates, the surface profile is reconstructed using the OpenCV image processing library, and the axis surface binary segmentation mask is stacked to reconstruct a three-dimensional segmented entity layer by layer. Finally, a three-dimensional segmentation result of the prostate organ region in the MR image is generated.

[0088] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A prostate contour segmentation method based on geometric intersection-union ratio loss, characterized in that, Includes the following steps: T2-weighted MR images were acquired, and the T2-weighted MR images were preprocessed to form an initial axial slice dataset. Determine whether the axial slices in the initial axial slice dataset contain the prostate region, and obtain a regression slice dataset containing the prostate region; The prostate region in the regression slice dataset is detected and located to obtain the coordinates of the prostate center point and the width and height of the detection box; Contour regression network model is constructed and trained. The contour regression network model consists of a shared weighted shared-MLP and a feature sequence module. The feature sequence module consists of one CMixer module without residual connections and seven continuous residual Res-CMixer modules. The CMixer module consists of a neighborhood feature aggregation module and a channel mixing sub-module. The neighborhood aggregation sub-module consists of a circular convolution module, and the channel mixing module consists of a Token Mixer structure module. An initial set of contour points is obtained based on the coordinates of the prostate center point and the width and height of the detection box, and a predicted set of contour points is obtained through a contour regression network model. Based on the predicted contour point set, a binary mask is reconstructed layer by layer to obtain the three-dimensional segmentation result of the prostate organ.

2. The prostate contour segmentation method based on geometric intersection-union ratio loss according to claim 1, characterized in that, The T2-weighted MR image preprocessing process includes: applying N4 offset field correction to the T2-weighted MR image, performing adaptive histogram equalization and then normalization, slicing the normalized MR image to obtain a continuous axial plane sequence image, performing offline data augmentation and dynamic data augmentation on the continuous axial plane sequence image to obtain an initial axial plane slice dataset, wherein the slices in the dataset are two-dimensional slices.

3. The prostate contour segmentation method based on geometric intersection-union ratio loss according to claim 1, characterized in that, The process of obtaining the regression slice dataset includes: using a classification network to identify whether the axial slices in the initial axial slice dataset contain the prostate region. The image input to the classification network is three adjacent slices in the axial plane. If the prostate region is contained, it is determined to be regression slice data; if not, it is removed without further processing.

4. The prostate contour segmentation method based on geometric intersection-union ratio loss according to claim 3, characterized in that, The detection and localization process includes: using the CenterNET architecture with the DLA network as the backbone to extract the network, using three consecutive slices from the regression slice dataset as input images for detection and localization, and obtaining the coordinates of the prostate target detection center point and the width and height of the detection and localization box.

5. The prostate contour segmentation method based on geometric intersection-union ratio loss according to claim 1, characterized in that, The process of constructing the initial contour point set includes: obtaining the midpoint coordinates of the four boundaries of the detection box by using the coordinates of the center point of the prostate target detection and the width and height data of the detection positioning box; correcting the four edge points of the prostate region based on the midpoint coordinates of the boundaries; constructing an octagonal contour in a clockwise order with the corrected edge points as the center; representing the octagonal contour in the form of a directed graph; and constructing the initial contour point set.

6. The prostate contour segmentation method based on geometric intersection-union ratio loss according to claim 1, characterized in that, The process of training the model includes: constructing a geometric reconstruction loss function, constraining the model training process based on the geometric reconstruction loss function, obtaining the area enclosed by the predicted polygonal contour and the gold standard contour, and using the maximization of the intersection area enclosed by the predicted polygonal contour and the gold standard contour as the optimization objective to constrain the model to update the weights. The geometric reconstruction loss function consists of a loss based on the geometric intersection-union ratio and a loss based on distance.

7. The prostate contour segmentation method based on geometric intersection-union ratio loss according to claim 6, characterized in that, The process of obtaining the predicted contour point set includes: extracting features from the regression slice dataset through the backbone network, inputting the feature sequence corresponding to the initial contour position into the contour regression network model, obtaining the coordinate offset of each point of the initial contour, adding it to the initial contour coordinates, and obtaining the predicted contour point set.

8. The prostate contour segmentation method based on geometric intersection-union ratio loss according to claim 1, characterized in that, The process of obtaining the three-dimensional segmentation results of the prostate organ includes: constructing the predicted contour point set into a boundary contour, filling the interior of the contour using computer vision technology, reconstructing the binary segmentation mask layer by layer, and stacking it to the original position to reconstruct the three-dimensional segmentation result.

Images