Brain tumor nuclear magnetic resonance image segmentation method based on deep learning

Abstract

The invention discloses a brain tumor nuclear magnetic resonance image segmentation method based on deep learning, and belongs to the technical field of medical image processing. According to the method, a multi-scale feature extraction module is introduced into a coding path and a decoding path of the U-Net network to enhance the extraction and fusion capability of the network on multi-scale information in a brain tumor nuclear magnetic resonance image; meanwhile, a coordinate attention module is introduced into the jump connection to enhance the modeling capability of the network on the dependency relationship between the tumor space position information and the channels, so that the common problems of detail loss, inaccurate small target segmentation and the like in the medical image segmentation task are solved. Besides, aiming at the serious category imbalance problem existing in the brain tumor data set, the method utilizes a mixed loss function to guide the training of the model, so that the model can better learn the characteristics of the multi-modal brain tumor in the training process.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a deep learning-based method for segmenting brain tumor MRI images. Background Technology

[0002] In recent years, the incidence of brain tumors, a very common brain disease, has been rising year by year. Based on their location of origin, brain tumors can be divided into two categories: primary brain tumors and secondary tumors. Among primary tumors, gliomas are the most common type, accounting for 50.1% of primary malignant tumors in adults. Brain tumors pose a significant threat to a patient's life and health, requiring early treatment. In clinical treatment, resection is the most typical treatment for gliomas. Although gliomas are relatively easy to detect, their diffuse and infiltrative growth in the brain, characterized by peritumoral edema and increased water content around the tumor, makes visually determining the tumor boundaries extremely difficult, both in medical imaging analysis and during surgery. Coupled with the complexity of the human brain structure, the resection of brain tumors faces formidable challenges, making accurate imaging of the brain and brain tumors crucial.

[0003] Magnetic Resonance Imaging (MRI) is a non-invasive imaging technique that aids in the clinical diagnosis of brain tumors. The resulting MRI images clearly show the location and extent of the tumor within the brain. In MRI, different modalities can be generated by adjusting the magnetic field and radiofrequency pulses in various ways. Different modalities reflect different pathological information, and combinations of multiple modalities are commonly used in diagnosis and treatment to comprehensively assess lesions.

[0004] In traditional medicine, brain tumor regions are typically segmented manually by clinicians based on brain MRI images. However, manual segmentation is time-consuming, and even highly experienced physicians' results are subject to subjectivity. Furthermore, with the increasing incidence of brain tumors and the growing number of patients, the location, size, and shape of tumors vary significantly among patients, making the development of segmentation templates extremely difficult. To improve the speed and accuracy of brain tumor medical image segmentation, automated brain tumor segmentation using computers has become an excellent option for the diagnosis and treatment of brain tumors.

[0005] In recent years, with the rapid development of deep learning, various deep learning algorithms have been widely applied to brain tumor segmentation tasks. By simulating the working principle of the human brain, deep learning can extract feature information from input images and construct deeper neural networks using multiple processing layers. Compared with traditional segmentation methods, deep learning improves efficiency, saves time and effort, and provides more accurate segmentation results. Therefore, deep learning has become the most popular and promising method in the field of medical image segmentation.

[0006] By utilizing efficient and accurate brain tumor medical image segmentation algorithms, automated segmentation of brain tumors can be achieved, which has a positive auxiliary role in the clinical diagnosis and subsequent treatment of brain tumor patients. Summary of the Invention

[0007] The purpose of this invention is to provide a deep learning-based method for segmenting brain tumor MRI images. This method aims to solve the problems of time-consuming, labor-intensive, and subjective manual segmentation in the current field of medical image analysis by achieving high-precision automated segmentation, thereby improving the utilization rate and processing efficiency of medical images in clinical practice.

[0008] The technical solution adopted in this invention is: a brain tumor MRI image segmentation method based on deep learning, which includes the following steps:

[0009] Step 1: Construct a segmentation model with a U-Net structure, which includes encoding and decoding branches;

[0010] The coding branch includes, in sequence, an initial coding block and several intermediate coding blocks. The initial coding block includes, in sequence, a first selective kernel convolutional layer and a multi-scale feature extraction module. The intermediate coding blocks include, in sequence, a second selective kernel convolutional layer and a multi-scale feature extraction module.

[0011] The decoding branch of the segmentation model includes an initial decoding block, several intermediate decoding blocks, and an encoded output block, wherein the number of intermediate decoding blocks is N-1, and N is the number of intermediate encoded blocks.

[0012] The initial decoded block is a third selective inner kernel convolutional layer used for upsampling, and the input of the initial decoded block is the input of the last intermediate encoded block;

[0013] The intermediate decoding block sequentially includes a connection layer, an MFE module, and a third selective inner kernel convolutional layer; for the 1st to the (N-1th)th encoding blocks, a coordinate attention module is introduced in the skip connection between any p-th decoding block and the Np-th decoding block, where p = 1, ..., N-1;

[0014] The encoded output block consists of a connection layer, an MFE module, and a selective inner kernel convolutional layer with a 1×1 kernel. A coordinate attention module is introduced in the skip connection between the encoded output block and the initial encoded block.

[0015] Step 2: Based on the collected brain tumor MRI image dataset, the constructed U-Net segmentation model is trained using deep learning. When the preset training convergence conditions are met (the number of training iterations or the loss function value reaches the desired condition), a segmentation model for brain tumor MRI image segmentation is obtained.

[0016] Furthermore, the first and second selective kernel convolutional layers have the same convolutional kernel, and the stride of the second selective kernel convolutional layer is set to 2.

[0017] Furthermore, the convolution kernels of the first, second, and third selective kernel convolutional layers are all set to 2×2.

[0018] Furthermore, the multi-scale feature extraction module includes several stacked selective kernel convolutional blocks and a dilated spatial pyramid pooling module; wherein, the skip connection branch between the input of the first selective kernel convolutional block and the output of the dilated spatial pyramid pooling module also includes a convolutional layer with a 1×1 kernel and a batch normalization layer; the selective kernel convolutional block includes a selective kernel convolutional layer, a batch normalization layer and an activation function layer in sequence.

[0019] Furthermore, the kernels of the selective kernel convolutional layers in the selective kernel convolutional blocks included in the multi-scale feature extraction module are all set to 3×3.

[0020] Furthermore, the multi-scale feature extraction module includes two selective kernel convolutional blocks.

[0021] Furthermore, in step 2, when training the segmentation model with the constructed U-Net structure using deep learning, a hybrid loss function is employed, the expression of which is:

[0022] Loss=λ G L G +λ C L C

[0023] Where Loss is the mixed loss function, L G For the generalized Dice loss function, L C Let λ be the cross-entropy loss function; G , λ C These are the weighting coefficients used to balance the two loss functions.

[0024] The technical solution provided by this invention brings at least the following beneficial effects:

[0025] This invention introduces a multi-scale feature extraction module into the encoding and decoding paths of the U-Net network to enhance the network's ability to extract and fuse multi-scale information from brain tumor MRI images. Simultaneously, a coordinate attention module is introduced into the skip connections to enhance the network's ability to model tumor spatial location information and inter-channel dependencies, thereby solving common problems in medical image segmentation tasks such as loss of detail and inaccurate segmentation of small objects. Furthermore, addressing the severe class imbalance problem in brain tumor datasets, this invention utilizes a hybrid loss function to guide model training, enabling the model to better learn the features of multimodal brain tumors during training. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 Diagram of brain tumor segmentation network structure;

[0028] Figure 2 Here is a structural diagram of the multi-scale feature extraction module;

[0029] Figure 3 Diagram of the pyramid pooling module structure for hollow spaces;

[0030] Figure 4 Here is a diagram of the coordinate attention module structure;

[0031] Figure 5 Example image of the segmentation result. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present invention.

[0033] Existing U-Net-based methods for brain tumor MRI image segmentation cannot effectively extract and fuse multi-scale feature information from brain tumor MRI images to improve segmentation performance. To enhance the network's ability to perceive and extract multi-scale features, this invention designs and introduces a multi-scale feature extraction module to enhance the feature extraction capability of the original model. Simultaneously, a coordinate attention module is introduced for explicit coordinate modeling and channel-space joint optimization to address the position sensitivity issue in medical image segmentation.

[0034] In this embodiment of the invention, during the segmentation of brain tumor MRI images, a multi-scale feature extraction module is introduced into the encoding and decoding paths based on the original U-Net network architecture to enhance the network's ability to extract and fuse multi-scale information in brain tumor MRI images. Simultaneously, a coordinate attention module is introduced into the skip connections to enhance the network's ability to model tumor spatial location information and inter-channel dependencies, thereby solving common problems in medical image segmentation tasks such as detail loss and inaccurate segmentation of small targets. Furthermore, to address the severe class imbalance problem present in brain tumor datasets, a hybrid loss function is used to guide model training, enabling the model to better learn the features of multimodal brain tumors during training.

[0035] In this embodiment of the invention, a multi-scale feature extraction module is introduced into the encoding / decoding path and the bottleneck layer to enhance the model's ability to extract multi-scale features and avoid missegmentation caused by getting stuck in local features.

[0036] In one embodiment, the multi-scale feature extraction module employs a residual structure based on selective kernel convolution and dilated spatial pyramid pooling. The residual structure mitigates the gradient vanishing problem that arises as the network deepens, enhancing the network's feature extraction capability. Replacing traditional convolutional layers in U-Net with selective kernel convolution dynamically selects convolutional kernels of different sizes based on the input feature map, thereby capturing multi-scale features and improving the model's expressive power. A dilated spatial pyramid pooling module is added after two selective kernel convolutions. By employing dilated convolutions with different dilation rates, the receptive field is expanded while maintaining high-resolution feature maps, enhancing the model's ability to process multi-scale detail information. Furthermore, since consecutive max-pooling operations in the encoding path can degrade image resolution, the max-pooling operations in the backbone network are replaced with selective kernel convolutions with a stride of 2.

[0037] In this embodiment of the invention, a coordinate attention module is added to the skip connection to combine channel attention with spatial location information. By decomposing channel attention into a one-to-one encoding process of one-dimensional features, feature aggregation in two different spatial directions is achieved, thereby more accurately locating the tumor target.

[0038] Furthermore, the loss function used during model (U-Net segmentation model) training is a hybrid loss function, the expression of which is:

[0039] Loss=λ G L G +λ C L C

[0040] In the formula, L G For generalized Dice loss; L C For cross-entropy loss; λ G , λ C These are the weighting coefficients used to balance the two loss functions.

[0041] Specifically, the generalized Dice loss and cross-entropy loss are expressed as follows:

[0042] (1) Generalized Dice loss

[0043]

[0044] In the formula, C represents the total number of tumor category labels; ω j gt represents the weight of the j-th category, typically taken as the reciprocal frequency of the category; N is the total number of pixels; ij Let seg be the true value of the j-th class label at the i-th pixel. ij Let be the predicted value of the j-th class label at the i-th pixel.

[0045] (2) Cross-entropy loss

[0046]

[0047] In the formula, C is the total number of tumor category labels, N is the total number of pixels, and gt ij Let seg be the true value of the j-th class label at the i-th pixel. ij Let be the predicted value of the j-th class label at the i-th pixel.

[0048] The generalized Dice loss function considers the weight of each class, avoiding the excessive influence of certain classes on the loss function when the sample size is small. It is particularly effective in handling class imbalance, preventing the model from being overly biased towards classes with higher frequency; cross-entropy can measure the distance between the probability distribution predicted by the model and the true label distribution.

[0049] In one embodiment, the present invention employs as follows: Figure 1The segmentation model with the U-Net structure shown has the following encoding branches: an initial encoding block and several intermediate encoding blocks. The initial encoding block consists of a first selective kernel convolutional layer and a multi-scale feature extraction (MFE) module. The intermediate encoding blocks consist of a second selective kernel convolutional layer and an MFE module. The first and second selective kernel convolutional layers have the same convolutional kernel, but the stride of the second selective kernel convolutional layer is set to 2. The decoding branch of the segmentation model consists of an initial decoding block, several intermediate decoding blocks (the number of which is one less than the number of intermediate encoding blocks), and an encoding output block. The initial decoding block is a selective kernel convolutional layer (or third selective kernel convolutional layer) with a stride of 2 used for upsampling. The input of the initial decoding block is the input of the last intermediate encoding block. The intermediate decoding blocks consist of a connection layer, an MFE module, and a third selective kernel convolutional layer. For the 1st to N-1th encoding blocks, coordinate attention is introduced in the skip connection between the p-th (p = 1, ..., N-1)th decoding block and the Np-th decoding block. The Attention (CA) module, the encoded output block includes a connection layer, an MFE module and a fourth selective inner kernel convolutional layer (with a 1×1 kernel), and the CA module is also introduced in the skip connection between the encoded output block and the initial encoded block.

[0050] In one embodiment, firstly, a 160×160×4 four-channel brain tumor image is input into the model. After passing through a selective kernel convolutional layer with a kernel size of 2×2, the image enters the multi-scale feature extraction module. The specific structure of the multi-scale feature extraction module is as follows: Figure 2 As shown, it includes multiple stacked selective kernel convolutional blocks (selective kernel convolutional layers, batch normalization, and activation functions) and atrous spatial pyramid pooling (ASPP). The skip connections (between input and output) of the MFE module also include convolutional layers with 1×1 kernels and batch normalization. This embodiment includes two selective kernel convolutional blocks; that is, firstly, the input feature map passes through two 3×3 selective kernel convolutional layers (SK-Conv), with batch normalization and ReLU activation performed after each convolution. SK-Conv mainly implements SK convolution through three operators: Split, Fuse, and Select. The splitting stage modifies the input feature map... Multi-branch convolution operations are performed. C represents the number of channels, and W×H represents the width and height of the feature map. The multi-branch convolution operation expands the receptive field size exponentially based on a 3×3 kernel, successively expanding it to 3×3, 5×5, and 7×7. Each of the three branches uses a different kernel size to generate its corresponding feature map, capturing features at different scales. In the fusion stage, the feature maps from the three branches obtained in the separation stage are summed to obtain U, which is then processed by global average pooling to obtain the channel descriptors. C′ represents the number of feature map channels after global average pooling. Next, a weight vector S is generated through a fully connected layer; during the selection phase, the selection weights are normalized using the Softmax function to obtain the weights a for each branch. i The feature maps of each branch are then weighted and summed according to their respective weights to obtain the final output V.

[0051] Then, the obtained feature map is subjected to hollow spatial pyramid pooling, the structure of which is as follows: Figure 3 As shown, in the global average pooling branch, global average pooling is used to compress the global information of the entire feature map into a 1×1 feature vector, and then a 1×1 convolution is used to upsample it back to its original size. In the dilated convolution branch, three parallel 3×3 dilated convolutions with dilation rates r set to 2, 4, and 6 are used to learn features for three tumor regions: the necrotic tumor core, the enhanced tumor region, and the peritumoral edema region (tumor-invading tissue), respectively. In the 1×1 convolution branch, 1×1 convolution is used to compress dimensions and adjust the number of channels.

[0052] y = Conv 1×1 (Concat(C 1,1 (x),C 3,2 (x),C 3,4 (x),C 3,6 (x),Up(GAP(x))))

[0053] In this code, Conv() represents a convolution operation, with the subscript indicating the kernel; Concat() represents a concatenation operation along the channel dimension; Up() represents an upsampling operation; GAP() represents a global average pooling operation; x represents the input feature map; and C() represents a dilated convolution operation, with its two subscripts indicating the kernel and dilation rate, respectively. For example, C... 3,2 (x) represents a dilated convolution with a dilation rate of 2 and a kernel size of 3×3 on x.

[0054] Meanwhile, in the residual structure, a single 1×1 ordinary convolution is used to adjust the number of channels and perform batch normalization.

[0055] The two feature maps obtained from the above operations are added together and then fed into the next encoding layer and the spatial attention module in the skip connection, respectively. Specifically, the output feature map of the current encoding layer, after being fed into the next encoding layer, undergoes pooling using a 2×2 selective kernel convolutional layer with a stride of 2.

[0056] After the feature map is input to the skip connection, it passes through a coordinate attention module, the structure of which is as follows: Figure 4 As shown. The CA module encodes channel relationships and long-range dependencies using precise location information, mainly including two steps: coordinate information embedding and coordinate attention generation. Coordinate information embedding decomposes the global average pooling operation into a one-to-one encoding operation of one-dimensional features, and then encodes the input feature map... Pooling is performed along the X and Y directions using two pooling kernels of different sizes, H×1 and 1×W, respectively, generating feature maps of size C×H×1 and C×1×W. and The coordinate attention generation transforms the C×1×W feature map, performs a concatenation operation with the C×H×1 feature map, and then uses a shared 1×1 convolution for dimensionality reduction to obtain intermediate feature maps of spatial information in the X and Y directions. Divide f into two separate tensors along the spatial dimension. and Then, two 1×1 convolutions are used to combine the feature map f h and f w Align the input feature map with the number of channels, pass it through the sigmoid activation function to obtain the attention vector, and then input it into the corresponding layer decoder.

[0057] On the decoding path, for the output feature map of the next layer, after performing a concat operation with the output feature map of the corresponding layer encoder passed in by the skip connection, it enters the multi-scale feature extraction module again, and then performs upsampling through a 2×2 selective kernel convolution with a stride of 2, and enters the next layer decoder. Each decoder consists of a splicing layer and an MFE module.

[0058] A 1×1 selective kernel convolution is performed on the output feature map of the multi-scale feature extraction module in the last layer decoder to obtain the final segmented image.

[0059] After the network was built, Adam was selected as the optimizer during network training. Dropout regularization was employed to prevent overfitting. Cosine annealing was used as the learning rate scheduling strategy to help the model escape local optima and accelerate model convergence.

[0060] The network constructed and trained based on the above steps was used to test its segmentation performance using the BraTS2020 dataset. A benchmark network, U-Net, was added for comparison. The segmentation results are shown in the example below. Figure 5 As shown in Table 1, the segmentation results are as follows.

[0061] Table 1. Segmentation performance of the brain tumor MRI image segmentation model

[0062]

[0063] As shown in Table 1, the method proposed in this embodiment of the invention was used to segment MRI images of brain tumors. On the BraTS2020 dataset, the Dice coefficients for the WT, TC, and ET tumor regions reached 85.6%, 83.2%, and 79.4%, respectively, with an average Dice coefficient of 82.7%. Compared to the original U-Net, the method proposed in this embodiment of the invention improved the Dice coefficients for the WT, TC, and ET tumor regions by 6.73%, 5.45%, and 2.32%, respectively, with an average improvement of 4.82%. The method proposed in this embodiment of the invention achieves high-precision automated segmentation of brain tumors, which has a positive auxiliary role in the clinical diagnosis and treatment of brain tumor patients.

[0064] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0065] Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include at least one of those features.

[0066] Any process or method description described in this specification can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order according to the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0067] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0068] Note that, unless otherwise explicitly stated, all features disclosed in this specification can be replaced by alternative features for achieving the same, equivalent, or similar purpose. Therefore, unless explicitly stated otherwise, each disclosed feature is merely one example of a set of equivalent or similar features. Where used, "further," "preferably," "even further," and "more preferably" are simply starting points for describing another embodiment based on the foregoing embodiments, the combination of which with the foregoing embodiments constitutes the complete configuration of another embodiment. Any combination of several "further," "preferably," "even further," or "more preferably" settings following the same embodiment constitutes yet another embodiment.

[0069] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0070] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A deep learning-based method for segmenting brain tumor MRI images, characterized in that, Includes the following steps: Step 1: Construct a segmentation model with a U-Net structure, which includes encoding and decoding branches; The coding branch includes, in sequence, an initial coding block and several intermediate coding blocks. The initial coding block includes, in sequence, a first selective kernel convolutional layer and a multi-scale feature extraction module. The intermediate coding blocks include, in sequence, a second selective kernel convolutional layer and a multi-scale feature extraction module. The decoding branch of the segmentation model includes an initial decoding block, several intermediate decoding blocks, and an encoded output block, wherein the number of intermediate decoding blocks is N-1, and N is the number of intermediate encoded blocks. The initial decoded block is a third selective inner kernel convolutional layer used for upsampling, and the input of the initial decoded block is the input of the last intermediate encoded block; The intermediate decoding block sequentially includes a connection layer, an MFE module, and a third selective inner kernel convolutional layer; for the 1st to the (N-1th)th encoding blocks, a coordinate attention module is introduced in the skip connection between any p-th decoding block and the Np-th decoding block, where p = 1, ..., N-1; The encoded output block consists of a connection layer, an MFE module, and a selective inner kernel convolutional layer with a 1×1 kernel. A coordinate attention module is introduced in the skip connection between the encoded output block and the initial encoded block. Step 2: Based on the collected brain tumor MRI image dataset, the constructed U-Net segmentation model is trained using deep learning. When the preset training convergence condition is met, a segmentation model for brain tumor MRI image segmentation is obtained.

2. The method as described in claim 1, characterized in that, The first and second selective kernel convolutional layers have the same convolutional kernel, and the stride of the second selective kernel convolutional layer is set to 2.

3. The method as described in claim 1, characterized in that, The kernels of the first, second, and third selective kernel convolutional layers are all set to 2×2.

4. The method as described in claim 1, characterized in that, The multi-scale feature extraction module includes several stacked selective kernel convolutional blocks and a dilated spatial pyramid pooling module. The jump connection branch between the input of the first selective kernel convolutional block and the output of the dilated spatial pyramid pooling module also includes a convolutional layer with a 1×1 kernel and a batch normalization layer. The selective kernel convolutional block includes a selective kernel convolutional layer, a batch normalization layer and an activation function layer in sequence.

5. The method as described in claim 4, characterized in that, The multi-scale feature extraction module includes selective kernel convolutional blocks, where the kernels of the selective kernel convolutional layers are all set to 3×3.

6. The method as described in claim 4 or 5, characterized in that, The multi-scale feature extraction module includes two selective kernel convolutional blocks.

7. The method as described in claim 1, characterized in that, In step 2, when training the segmentation model with the constructed U-Net structure using deep learning, a hybrid loss function is employed, the expression of which is: Loss=λ G L G +λ C L C Where Loss is the mixed loss function, L G For the generalized Dice loss function, L C Let λ be the cross-entropy loss function; G , λ C These are the weighting coefficients used to balance the two loss functions.

Images