Automatic segmentation system and method based on multi-mode MRI brain tumor image

Abstract

The invention discloses an automatic segmentation system and method based on a multi-mode MRI brain tumor image, and the system comprises an MCFE-Net network composed of an encoder and a decoder, and achieves the automatic segmentation of lesion regions WT, ET and TC of the MRI brain tumor image through the MCFE-Net network. A multi-convolution module is arranged in the middle layer of the encoder and the decoder, a cavity space receptive field module is arranged in the deepest layer of the encoder, and a modal feature attention module is arranged at the jump joint of the encoder and the decoder; wherein the multi-convolution module is used for improving the coding capability of the network while improving the width of the network; the cavity space receptive field module is used for fully utilizing the deep enriched semantic features and providing multi-scale context information so as to enhance the extraction of the network and utilize the deep semantic features; and the modal feature attention module is used for eliminating the influence of noise and irrelevant backgrounds and enhancing local expression. According to the method, the WT, ET and TC regions in the MRI brain tumor image can be efficiently segmented.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to an automatic segmentation system and method based on multimodal MRI brain tumor images. Background Art

[0002] Glioma is the most common primary tumor of the central nervous system. Its high invasiveness and damage to brain tissue can have a serious impact on patients' lives and health. The most common brain tumors in adults are primary central nervous system lymphomas and gliomas, with gliomas accounting for over 80% of malignant tumors. Based on the aggressiveness of gliomas, they are broadly divided into high-grade gliomas (HGG) and low-grade gliomas (LGG). HGG is a malignant tumor that grows rapidly and easily forms abnormal tissue. HGG patients have a high mortality rate, with a five-year survival rate of only 5.5%. LGG can be benign or malignant, grows more slowly, and may recur after treatment and develop into HGG. Therefore, timely diagnosis and treatment of brain tumors is of great significance.

[0003] With the development of medical imaging, imaging technology plays an important role in the treatment evaluation of brain tumor patients, providing doctors with a clear view of the internal structure of the human body. Common imaging technologies in the field of brain medicine include computed tomography, single-photon emission computed tomography, positron emission tomography, magnetic resonance spectroscopy, and magnetic resonance imaging. These imaging technologies can provide information on the shape, size, and location of brain tumors, facilitating accurate diagnosis and treatment planning. However, because MRI is a non-invasive imaging modality, it has unique advantages such as good soft tissue contrast, no radiation damage, no bone artifacts, and the ability to image multiple parameters. Therefore, MRI is particularly suitable for clinical brain lesion examinations.

[0004] In recent years, artificial intelligence and deep learning technologies have developed rapidly, leading to significant advances in the data processing and generalization capabilities of neural networks. Brain tumor image segmentation methods that combine brain tumor image analysis with deep learning have effectively improved the quality of brain tumor image segmentation, reduced physician workload, and increased diagnostic efficiency. Currently, deep learning is the optimal approach for brain tumor segmentation. With the emergence of convolutional neural networks (CNNs), thanks to their end-to-end learning mechanism, they extract deep image features and use them to segment target regions. Numerous segmentation algorithms have subsequently been proposed based on these features. Among these, the UNet algorithm has been the most widely adopted algorithm in the field of brain tumor segmentation since its introduction. Its simple structure and high efficiency have formed the foundation of numerous subsequent segmentation algorithms. Although deep learning-based brain tumor segmentation algorithms have achieved impressive performance, they still face numerous challenges. For example, current brain tumor segmentation models suffer from blurred boundary information and incomplete segmentation of small tumors, which remain to be addressed, and there is still room for improvement in model accuracy and robustness. SUMMARY

[0005] The present application aims to provide an automatic segmentation system and method based on multi-modal MRI brain tumor images, to solve the technical problem of how to quickly and accurately segment the WT, TC and ET regions in the brain tumor image.

[0006] The present application is implemented by the following technical scheme: an automatic segmentation system based on multi-modal MRI brain tumor images, comprising a MCFE-Net network composed of an encoder and a decoder, realizing the automatic segmentation of the lesion regions WT, ET and TC of the MRI brain tumor image through the MCFE-Net network; a multiple convolution module is arranged in the middle layer of the encoder and the decoder, a hollow spatial receptive field module is arranged in the deepest layer of the encoder, and a modal feature attention module is arranged at the jump connection of the encoder and the decoder. Among them, the multiple convolution module is used to improve the network width and enhance the encoding ability of the network; The hollow spatial receptive field module is used to make full use of the rich semantic features in the deep layer and provide multi-scale context information to enhance the extraction and utilization of deep semantic features of the network; The modal feature attention module is used to eliminate the influence of noise and irrelevant background and enhance local expression.

[0007] Further, the multiple convolution module includes multiple different convolutions, and the outputs of larger convolution kernels are added element by element through addition operation, while the outputs of smaller convolution kernels are added to the outputs of larger convolution kernels as residual structure branches.

[0008] Further, the multiple convolution module includes three parts: The first part includes three different convolutions: 1x1 convolution, 3x3 convolution and 5x5 convolution; The second part includes a first adder, which adds the outputs of 3x3 convolution and 5x5 convolution element by element through addition operation, and increases the channel information amount one by one; The third part includes a second adder, which adds the output of 1x1 convolution as the output of the residual structure branch to the output of the second part, allowing the gradient to flow directly through the network to alleviate the problem of gradient disappearance, and enabling more stable gradient transmission during back propagation.

[0009] Further, the hollow spatial receptive field module includes a main stem and two branches, the main stem is composed of a deep separation hollow convolution group, the two branches are residual branch and global receptive field branch respectively, and finally output after GSE attention focusing semantic feature.

[0010] Furthermore, the data processing process of the GSE attention is as follows: After the operation, global average pooling is adopted for channel dependency, a gate mechanism consisting of double full connection plus ReLU is used, and Sigmoid activation is used to obtain weights, and finally the scale operation is used to reweight and obtain the weighted output.

[0011] Furthermore, the modal feature attention module has a dual-input structure, which is used for skip connections to connect deep and shallow layers and automatically learn to focus on target structures of different shapes and sizes.

[0012] Furthermore, the modal feature attention module includes a Grid att submodule, which convolves the current layer input and the deep layer input in the dual input by 1×1 respectively so that the module can extract the effective features of the input feature map through deep learning. After the add operation, the effective information of the feature map will be further increased without increasing the number of parameters. The number of channels is reduced through convolution, and finally the spatial feature weight is obtained through sigmoid activation.

[0013] Furthermore, the modal feature attention module also includes a Glob att submodule, which performs a concat operation on the current layer input and the deep layer input, connects them in the channel dimension, compresses them into a 1×1×C feature vector through a global average pooling operation, and processes them through the FC layer and the RL layer, and finally obtains the channel feature weight through the sigmoid activation function.

[0014] An automatic segmentation method based on multimodal MRI brain tumor images is implemented based on the above-mentioned automatic segmentation system based on multimodal MRI brain tumor images, comprising the following steps: S1: Data division and preprocessing: MRI images are randomly divided into training set, validation set and test set according to the preset ratio; S2: Model training and testing: import the training set and validation set into the MCFE-Net network for training and validation, evaluate the model performance through the test set, and incorporate the best-performing model into the segmentation system; S3: Segmentation system construction, which combines the slicing and resizing functions of .nii files, as well as two-dimensional segmentation and evaluation, into a complete brain tumor segmentation system. The brain tumor segmentation system can automatically segment the lesion area of ​​MRI brain tumor images into WT, ET, and TC.

[0015] Furthermore, step S2 is specifically as follows: importing the training set and validation set into the MCFE-Net network for training and validation, using the composite loss function to calculate the training loss, updating the parameters according to the loss of the training set, and saving the best model with reference to the results of the validation set; in the testing phase, using DSC, Hausdorff distance, intersection-over-union ratio, and sensitivity to evaluate the performance of the model, and incorporating the model with the best performance into the segmentation system.

[0016] The beneficial effects of the present invention are that the present invention can quickly complete the segmentation task of a specified case and automatically calculate the evaluation indicators for the doctor's reference. Thanks to the MCFE-Net network's effective segmentation and lightness, the present invention can run normally on an ordinary computer and has more practical application value in the clinical field. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0018] Figure 1 This is a system block diagram of the present invention; Figure 2 This is the schematic diagram of the multiple convolution module; Figure 3 This is the schematic diagram of the void space receptive field module; Figure 4 This is the GSE attention principle diagram; Figure 5 This is the schematic diagram of the modal feature attention module; Figure 6 Construct a flow chart for the segmentation system. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0020] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0021] The following embodiments of the present invention are described in detail with reference to the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments may be combined with each other.

[0022] The present invention proposes a CNN with an encoder-decoder structure to address the problems of blurred boundaries of MRI brain tumor images and misclassification of small-sized tumors in segmentation results. The CNN can be used for automatic segmentation of WT, ET, and TC in the lesion areas of MRI brain tumor images. On the one hand, most networks ignore the width of the network, and the final segmented tumor has blurred and erroneous boundaries and details. A network with a certain width has better effects in dealing with boundary information and local information. On the other hand, after being encoded by the encoder layer by layer, the deep coding output has the characteristics of being smaller than the initial size and having the largest number of channels. After continuous encoding, it has the richest semantic features. Failure to fully utilize them will significantly affect the final segmentation accuracy. In addition, although the current network pays attention to spatial information and channel information, there is a problem of paying too much attention to spatial or channel information and ignoring the criticality of the other information.

[0023] The present invention proposes three modules to improve the final brain tumor segmentation accuracy in response to the existing segmentation difficulties, and obtains an efficient segmentation model: a CNN structure network with multiple convolutional feature extraction, named MCFE-Net, which can efficiently segment the WT, ET, and TC regions in MRI brain tumor images. In previous brain tumor segmentation studies, segmentation networks such as U-Net still have the problem of missing small-sized tumors and blurred boundaries of segmented regions in the final segmentation results. After analysis, the main reasons may be: (1) ignoring the width of the network. The width of the network is related to the ability to focus on the key areas of the tumor - boundaries, details, etc., and an overly wide network will also lead to overfitting. (2) The loss of spatial information and channel information. Spatial information and channel information are key factors in brain tumor segmentation. Spatial information itself enables the neural network to explore the global context and model long-term dependencies. Channel information is related to the ability of the neural network to focus on key areas. (3) Insufficient utilization of deep semantic features: Semantic features are one of the most important elements in brain tumor segmentation algorithms. The network’s ability to extract semantic features is strongly correlated with the quality of the final segmentation results. However, the identity of the encoding layer causes the network to ignore the semantically enriched characteristics of deep encoding.

[0024] See also Figure 1 In order to solve the above problems, the MCFE-Net in this invention will be designed from the following aspects: By replacing part of the encoder with the multiple convolutional modules proposed in this design, the network width is moderately increased while also improving the network's encoding capabilities.

[0025] A hollow spatial receptive field module is proposed in the deepest layer of the encoding part to fully utilize the semantic features enriched in the deep layer and provide multi-scale context information to enhance the network's extraction and utilization of deep semantic features.

[0026] A designed modal feature attention module is introduced at the jump connection. While focusing on encoding the objective channel information and spatial information of each layer due to its different size and number of channels, it eliminates the influence of noise and irrelevant background and enhances local expression.

[0027] MRI brain tumor images vary greatly from image to image. Compared to the UNet series of networks that only use a single-size convolution encoding format, the multi-convolution module designs a parallel convolution structure with multiple different convolutions to widen the network width. Different convolution kernels are selected in different convolutions to expand the receptive field and have stronger encoding capabilities. The module structure is as follows: Figure 2 As shown in the figure, after the input data is processed, it is divided into three parts. In the first part, the input is passed to three different convolutions, where different convolution kernels (1×1, 3×3, and 5×5) are selected to expand the receptive field. In the second part, the outputs of the larger convolution kernels (3×3 and 5×5) are added element-by-element, increasing the information content of each channel. In the third part, the output of the 1×1 convolution kernel is used as the output of the residual structure branch and added to the output of the second part. This allows the gradient to flow directly through the network, thereby alleviating the problem of vanishing gradients and ensuring more stable gradient transmission during backpropagation.

[0028] In addition, in order to make the network have stronger encoding capabilities while avoiding overfitting due to too many parameters, the network ultimately only uses multiple convolutional modules in the second, third, and fourth layers of the encoding and decoding parts.

[0029] Semantic features, as the meaning and contextual information of data, are a high-level representation of data. In the brain tumor segmentation task, how the network can efficiently extract semantic features is the most concerned issue for all segmentation algorithms. After continuous encoding, as the size of the data becomes smaller and the number of channels continues to increase, the deep coding layer will have weaker spatial position information and stronger semantic feature information than the shallower layer. Therefore, the network will have the richest semantic features in the deepest layer of the coding layer. In response to this network characteristic, a void spatial receptive field module is proposed, which can generate continuous multi-scale receptive fields. The module structure is as follows Figure 3 shown.

[0030] The module structure consists of a trunk and two branches. The trunk is mainly composed of a depth-separated dilated convolution group, and the two branches are the residual branch and the global receptive field branch. Finally, through the concat operation, the GSE attention focuses on the semantic features and outputs them. GSE attention can explicitly model the interdependence between convolutional feature channels to improve the representation ability of the network, as shown below. Figure 4 As shown in the figure. After the deep feature map passes through the parallel deep separation dilated convolution group and residual branch and global receptive field branch of the dilated spatial receptive field module, a concat operation is performed to associate the feature image and make it continuous, and the number of channels is increased several times. Due to the small size and multi-channel characteristics of the deep feature map, it has too many channels and is prone to overfitting. After reducing the number of channels by using deep convolution with less computation and parameters, After the operation, global average pooling is adopted for channel dependency, and the gate mechanism consisting of double full connection plus ReLU is used, and the weight is obtained by Sigmoid activation. Finally, the weight is re-weighted through the scale operation to obtain the weighted output.

[0031] The modal feature attention module has a dual input structure and is used for skip connections to connect deep and shallow layers, automatically learning to focus on target structures of different shapes and sizes, as shown below: Figure 5 As shown in the figure, due to the difference in feature map size and number of channels at each layer during network coding, deep and shallow features have different weights on spatial information and channel information. The modal feature attention module is used at the skip connection to associate deep and shallow feature maps while eliminating the influence of noise and irrelevant background, thereby enhancing local expression.

[0032] The modal feature attention module specifically includes the Grid att submodule and the Glob att submodule. In the Gridatt submodule, the current layer input and the deep layer input in the dual input are each subjected to a 1×1 convolution, enabling the module to extract effective features from the input feature map through deep learning. The add operation further increases the effective information of the feature map without increasing the number of parameters. The number of channels is reduced through convolution, and finally, spatial feature weights are obtained through sigmoid activation. In the Glob att submodule, the current layer input and the deep layer input are concat-connected in the channel dimension and compressed into a 1×1×C feature vector through a global average pooling operation. To learn the nonlinear interactions and non-mutually exclusive relationships between channels, they are processed through the FC layer and the RL layer, and finally, the channel feature weights are obtained through the sigmoid activation function.

[0033] Spatial information and channel information are important factors affecting the correct segmentation of small-sized tumors. Because the output size and channel of each layer of jump connection are different after encoding, the spatial information and channel information are also different. The modal feature receptive field module can automatically focus on the spatial information and channel information of the feature map, obtain spatial weights and channel weights, and perform information extraction to ultimately correctly segment small-sized tumors.

[0034] See also Figure 6 , the process of building a segmentation system is: Data segmentation and preprocessing. The data comes from the brain tumor segmentation dataset provided by the International Association for Medical Image Computing and Computer-Assisted Intervention Brain Tumor Segmentation Challenge. First, the 285 MRI images provided by the competition were randomly divided into training, validation, and test sets in an 8:1:1 ratio. Each image was then sliced ​​and resized to 160×160. Each image slice contains two conditions: a) an image with a tumor and a corresponding binary label; b) an image without a tumor and a corresponding pure black background. Therefore, the model must first determine whether the image contains a tumor region and then segment the tumor-containing area.

[0035] Model training and testing. The training and validation sets were imported into MCFE-Net for training and validation, using a composite loss function (dice loss and cross-entropy loss) to calculate the training loss. The model updated its parameters based on the training set loss, and the best model was saved based on the validation set results. Because the dice loss-guided training process can be unreliable, this project used the dice similarity coefficient (DSC) of the validation set as a reference. An early stopping mechanism monitored the change in DSC over a 20-epoch range. If the validation DSC did not increase within 20 epochs, the model was considered to have converged to the ideal range. Finally, during the testing phase, DSC, Hausdorff distance (HD), intersection of union (IOU), and sensitivity (Sens) were used to evaluate model performance, and the best-performing model was incorporated into the segmentation system.

[0036] Segmentation system construction. We integrated the following two functions into a complete workflow: (1) .nii file slicing and resizing, and (2) 2D segmentation and evaluation. We used PyQt5 to create a user interface, drawing function buttons and display boxes, and connecting them to corresponding callback procedures. This ultimately resulted in a simple brain tumor segmentation system.

[0037] To validate the effectiveness of the proposed MCFE-Net, we conducted ablation experiments with the backbone network and the addition and removal of various modules. We used a test set partitioned from the BraTS2019 dataset, retaining only sections and labels containing tumors. The test set results, shown in Table 1, show significant performance improvements for MCFE-Net on the test set, demonstrating its significant performance advantages on both datasets.

[0038] Table 1. Performance comparison of MSBP-Net and other networks

[0039] Furthermore, to facilitate clinicians' use, this paper uses PyQt5 to package the complete MRI brain tumor segmentation process into a segmentation system, consisting of two stages: 1. Slicing and 2. Segmentation. The physician simply selects the corresponding path in the operation box and clicks the corresponding function button to complete the segmentation of the tumor region in the MRI image. The system also automatically calculates the tumor segmentation accuracy for the physician's reference, making the entire process concise and clear, saving time and effort.

[0040] The present invention enhances the ability of the encoder to capture tumors. In the encoder part, compared with the weaker encoding ability of the convolutional coding layer of the U-Net series network, MCFE-Net uses multiple convolution modules to replace the encoder and decoder parts, and the parallel convolution group of the multiple convolution modules expands the width of the network, can provide a multi-level receptive field, and make the network have stronger encoding and decoding capabilities. U-Net ignores the characteristic that the deep layer has enriched semantic features after continuous encoding of the network, while MCFE-Net uses a void space receptive field module in the deep layer of the network. The void space receptive field module consists of three branches, which can generate multi-scale receptive fields, fully extract deep semantic features, and finally accurately complete boundary segmentation. The present invention can not only automatically segment the tumor area, but also automatically calculate according to the evaluation indicators, which can be used as a reference for doctors. The whole process is simple and easy.

[0041] For the sake of simplicity, the aforementioned embodiments are described as a series of actions. However, those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are preferred embodiments, and the actions involved are not necessarily required by this application.

[0042] The above embodiments describe 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 above embodiments and descriptions are merely illustrative of the principles of the present invention. Without departing from the spirit and scope of the present invention, modifications and variations made by those skilled in the art without departing from the spirit and scope of the present invention should be within the scope of protection of the appended claims.

Claims

1. An automatic segmentation system for brain tumor images based on multimodal MRI, characterized by: The proposed method includes an MCFE-Net network consisting of an encoder and a decoder, which automatically segment the WT, ET, and TC lesion areas of MRI brain tumor images. The network includes multiple convolutional modules in the middle layers of the encoder and decoder, a dilated spatial receptive field module in the deepest layer of the encoder, and a modal feature attention module at the jump connection between the encoder and decoder. Among them, multiple convolution modules are used to increase the network width and improve the network's encoding capabilities; The hollow spatial receptive field module is used to fully utilize the semantic features enriched in the deep layer and provide multi-scale context information to enhance the network's extraction and utilization of deep semantic features; The modal feature attention module is used to eliminate the influence of noise and irrelevant background and enhance local expression.

2. The automatic segmentation system based on multimodal MRI brain tumor images according to claim 1, characterized in that: The multiple convolution module includes multiple different convolutions, and adds the output of the larger convolution kernel element by element through an addition operation, while the output of the smaller convolution kernel is used as the output in the residual structure branch and added to the output result of the larger convolution kernel.

3. The automatic segmentation system based on multimodal MRI brain tumor images according to claim 2, characterized in that: The multi-convolution module consists of three parts: The first part includes three different convolutions: 1×1 convolution, 3×3 convolution, and 5×5 convolution; The second part, including the first adder, adds the outputs of the 3×3 convolution and the 5×5 convolution element by element, increasing the amount of information of each channel one by one; The third part, including the second adder, adds the output of the 1×1 convolution as the output of the residual structure branch to the output of the second part, allowing the gradient to flow directly through the network to alleviate the problem of gradient disappearance and enable more stable gradient transmission during backpropagation.

4. The automatic segmentation system based on multimodal MRI brain tumor images according to claim 1, characterized in that: The dilated spatial receptive field module consists of a trunk and two branches. The trunk consists of a depth-separated dilated convolution group, and the two branches are the residual branch and the global receptive field branch. Finally, the concat operation is performed and the GSE attention focuses on the semantic features before output.

5. The automatic segmentation system based on multimodal MRI brain tumor images according to claim 4, characterized in that: The data processing process of the GSE attention is as follows: After the operation, global average pooling is adopted for channel dependency, a gate mechanism consisting of double full connection plus ReLU is used, and Sigmoid activation is used to obtain weights, and finally the scale operation is used to reweight and obtain the weighted output.

6. The automatic segmentation system based on multimodal MRI brain tumor images according to claim 1, characterized in that: The modal feature attention module has a dual-input structure, which is used for skip connections to connect deep and shallow layers and automatically learn to focus on target structures of different shapes and sizes.

7. The automatic segmentation system based on multimodal MRI brain tumor images according to claim 6, characterized in that: The modal feature attention module includes a Grid att submodule, which convolves the current layer input and the deep layer input in the dual input by 1×1 respectively so that the module can extract the effective features of the input feature map through deep learning. After the add operation, the effective information of the feature map will be further increased without increasing the number of parameters. The number of channels is reduced through convolution, and finally the spatial feature weight is obtained through sigmoid activation.

8. The automatic segmentation system based on multimodal MRI brain tumor images according to claim 7, characterized in that: The modal feature attention module also includes a Glob att submodule, which performs a concat operation on the current layer input and the deep layer input, connects them in the channel dimension, compresses them into a 1×1×C feature vector through a global average pooling operation, and processes them through the FC layer and the RL layer, and finally obtains the channel feature weight through the sigmoid activation function.

9. A method for automatic segmentation of brain tumor images based on multimodal MRI, implemented based on the automatic segmentation system for brain tumor images based on multimodal MRI according to any one of claims 1 to 8, characterized in that: The steps include: S1: Data division and preprocessing: MRI images are randomly divided into training set, validation set and test set according to the preset ratio; S2: Model training and testing: import the training set and validation set into the MCFE-Net network for training and validation, evaluate the model performance through the test set, and incorporate the best-performing model into the segmentation system; S3: Segmentation system construction, which combines the slicing and resizing functions of .nii files, as well as two-dimensional segmentation and evaluation, into a complete brain tumor segmentation system. The brain tumor segmentation system can automatically segment the lesion area of ​​MRI brain tumor images into WT, ET, and TC.

10. The automatic segmentation method based on multimodal MRI brain tumor images according to claim 9, characterized in that: Step S2 is as follows: the training set and validation set are imported into the MCFE-Net network for training and validation, the training loss is calculated using a composite loss function, the parameters are updated according to the loss of the training set, and the best model is saved with reference to the results of the validation set; in the testing phase, the performance of the model is evaluated using DSC, Hausdorff distance, intersection-over-union ratio, and sensitivity, and the model with the best performance is incorporated into the segmentation system.

Images