Cervical cancer clinical target image segmentation method

Abstract

The invention provides a cervical cancer clinical target region image segmentation method. The method comprises the following steps: data preprocessing; a Dense U-Net neural network model is constructed; a Dense U-Net neural network model is trained; segmenting a cervical cancer clinical target region image and the like. The invention discloses a cervical cancer clinical target image segmentation method based on a Dense U-Net neural network, and the method comprises the steps: providing a Dense U-net network fusing dense connection and multi-scale feature fusion, and improving the parameter efficiency through a feature reuse mechanism in a dense block; multi-scale features are reserved through cross-layer cascading jump connection; optimizing small target segmentation precision by a mixed loss function; and the lightweight design realizes clinical deployment feasibility.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of cervical cancer clinical target volume image segmentation methods, in particular to a cervical cancer clinical target volume image segmentation method based on a Dense U-Net neural network. BACKGROUND

[0002] Cervical cancer is one of the common malignant tumors in women, and radiotherapy is an important treatment for cervical cancer. In the process of radiotherapy, accurate segmentation of the clinical target volume (CTV) of cervical cancer is crucial for developing a reasonable radiotherapy plan, improving treatment effectiveness, and reducing damage to normal tissues.

[0003] Traditional cervical cancer clinical target volume segmentation methods mainly rely on manual delineation by doctors, which is time-consuming and labor-intensive, and the delineation results of different doctors have great subjectivity and variability, making it difficult to ensure consistency and accuracy of segmentation. With the development of computer technology, automatic or semi-automatic segmentation methods based on image processing and machine learning have gradually emerged. Patent CN 119810118 A discloses an anatomical prior position information guided cervical cancer clinical target volume segmentation method, which uses CT image data and a reference CTV of a cervical cancer patient as basic data, calculates a corresponding set of distance information image data Dimg based on the basic data, constructs a multi-task dual-branch feature fusion network MTF-Net based on UNet network, uses the reference CTV and Dimg as a reference for training MTF-Net, and uses the CT image data as input information for intelligent segmentation of MTF-Net. The reference CTV refers to DICOM format data of cervical cancer CTV manually delineated and determined according to professional guidelines and processes. The distance information image data Dimg refers to distance information between surrounding adjacent organs and cervical cancer CTV.

[0004] However, existing methods still have some problems, for example: traditional Convolutional Neural Network (CNN) has limitations in network structure when processing medical images, making it difficult to fully extract context information and detailed features of the image, resulting in low segmentation accuracy; U-Net network has achieved good results in medical image segmentation, but there is some information loss in the information transmission process, and for cervical cancer target area with complex boundary and variable shape, the segmentation effect still needs to be improved.

[0005] U-Net is a convolutional neural network (CNN) architecture proposed by researchers Olaf Ronneberger, Philipp Fischer, and Thomas Brox in 2015, mainly used for biomedical image segmentation. It consists of symmetric search paths (shrinkage paths) and expansion paths, with features such as skip connections. Dense U-Net is an improved network based on U-Net combined with the idea of DenseNet. The core of DenseNet is to introduce dense connections in the network, making each layer connected to all previous layers, promoting feature propagation and reuse, and increasing the diversity of features in the network.

[0006] The effect of changing the original U-Net simple structure to DenseNet type can increase the hyperparameters in the network. In U-Net, as the network depth increases, there may be problems such as feature loss, while Dense U-Net can better reuse features from different layers through dense connections, each layer can get additional input from the previous layer, thus improving the network's expression ability, which helps to more accurately capture feature information in the image, such as more accurately segmenting cells, tumors, and other targets in medical image segmentation.

[0007] U-Net has fewer layers and parameters compared to many current networks, and is prone to overfitting during training. Dense U-Net, due to its dense connection characteristics, can increase the complexity and depth of the network without significantly increasing the number of parameters, while the feature reuse brought by dense connections can more effectively utilize data, thus alleviating the overfitting problem and improving the model's generalization ability on different datasets.

[0008] Due to the excellent performance of U-Net in medical image segmentation, Dense U-Net has wide applications in the medical image field. Medical data acquisition costs are high, and samples are relatively small. Dense U-Net can effectively train on limited samples, and its enhanced feature expression ability may have better results for medical image segmentation with multiple cell types and complex staining conditions. At the same time, in other fields that require image segmentation, such as remote sensing image segmentation and target detection, Dense U-Net may also play an important role, providing support for subsequent analysis and decision-making through accurate feature extraction and segmentation. SUMMARY

[0009] The purpose of the present application is to provide a cervical cancer clinical target volume image segmentation method, which is a cervical cancer clinical target volume image segmentation method based on a DenseU-Net neural network, a Dense U-net network that fuses Dense Connection and multi-scale feature fusion is proposed, which is achieved by: 1, the feature reuse mechanism in the dense block improves the parameter efficiency; 2, the cross-layer cascade jump connection retains the multi-scale feature; 3, the mixed loss function optimizes the small target segmentation accuracy; 4, the lightweight design realizes the feasibility of clinical deployment.

[0010] The technical scheme adopted by the present application to achieve the technical purpose is: a cervical cancer clinical target volume image segmentation method, comprising the following steps:

[0011] Step 1, data preprocessing: collect CT images of cervical cancer patients, and perform normalization processing on the images; manually label the clinical target volume to generate the corresponding labeled image, and divide the original image and the labeled image into a training set, a validation set and a test set;

[0012] Step 2, constructing a Dense U-Net neural network model;

[0013] Step 3, training the Dense U-Net neural network model;

[0014] Step 4, cervical cancer clinical target volume image segmentation; input the original image into the trained Dense U-Net neural network model, and obtain the output result of the model through forward propagation, the output result is a segmentation image with the same size as the input image, each pixel point in the segmentation image corresponds to a class label, which is used to indicate whether the pixel point belongs to the cervical cancer clinical target volume, thereby realizing the segmentation of the cervical cancer clinical target volume image.

[0015] 3. Further, in the cervical cancer clinical target volume image segmentation method described above: in step 1, when normalizing the image: the pixel value of the image is normalized to the range of [0, 1] through the formula:

[0016]

[0017] where (x) is the original pixel value, x min and x max are the minimum and maximum pixel values in the image, respectively.

[0018] Further, in the cervical cancer clinical target volume image segmentation method described above: in step 1, while normalizing the image, the image is cropped and scaled to a size of 256x256.

[0019] Further, in the cervical cancer clinical target volume image segmentation method, a tumor radiotherapy physician with rich experience manually labels the clinical target volume to generate a labeled image.

[0020] Further, in the cervical cancer clinical target volume image segmentation method, the 100 cases of data are further divided into a training set, a validation set and a test set according to a ratio of 7:1:2.

[0021] Further, in the cervical cancer clinical target volume image segmentation method, in the step 2, the encoder part comprises four dense blocks, each of which comprises four convolutional layers, and the convolutional kernel size is 3*3, the step is 1 and the padding is 1.

[0022] The Dense U-Net neural network model comprises an encoder part, a decoder part and a skip connection part.

[0023] The encoder part comprises at least one dense block and a transition layer for reducing the resolution of a feature map and reducing the amount of calculation, and each dense block is composed of at least two convolutional layers, and the input of each convolutional layer in the dense block is the splicing of the outputs of all previous convolutional layers.

[0024] The decoder part is used for upsampling a feature map to restore the resolution of an image, and is symmetrical to the encoder part and comprises at least one dense block and a transition layer.

[0025] The skip connection part splices the feature maps of the corresponding layers in the encoder and the feature maps of the corresponding layers in the decoder.

[0026] Further, in the cervical cancer clinical target volume image segmentation method, in the step 2, the encoder part comprises four dense blocks, each of which comprises four convolutional layers, and the convolutional kernel size is 3*3, the step is 1 and the padding is 1.

[0027] Further, in the cervical cancer clinical target volume image segmentation method, in the step 3, the following steps are included.

[0028] In step 301, the original image of the training set is input into the constructed Dense U-Net neural network model, and the output result of the model is calculated through forward propagation.

[0029] In step 302, the loss function is calculated according to the output result of the model and the labeled image.

[0030] In step 303, the parameters of the network are updated through the back propagation algorithm, and the weights and biases of the network are continuously adjusted to minimize the loss function.

[0031] In the training process, the model is verified using the verification set, the training parameters are adjusted according to the results of the verification set, when the loss function on the verification set no longer decreases or reaches the preset training times, the training is stopped, and the trained DenseU-Net neural network model is obtained.

[0032] Further, in the cervical cancer clinical target volume image segmentation method described above: in the step 3 training process: the loss function (L) adopts a weighted sum of cross-entropy loss function L ce and Dice coefficient loss function L dice ; the Adam optimizer is used to update the network parameters, the initial learning rate is set to 0.001, and the batch size is 8; every 5 epochs, the model is verified using the verification set, the parameters including the learning rate are adjusted according to the loss function and the segmentation accuracy on the verification set; when the loss function on the verification set does not decrease continuously for 10 epochs, the training is stopped.

[0033] The Dense U-Net neural network model is adopted, the feature information is fully utilized through the dense connection structure, the information loss is reduced, the jump connection is combined, the capturing ability of the network to the detail information is enhanced, the features of the cervical cancer clinical target volume can be more accurately extracted, and the segmentation accuracy is improved.

[0034] The cross-entropy loss function and the Dice coefficient loss function are combined to serve as the loss function, the class imbalance problem can be better balanced, and the segmentation performance of the model to small targets (such as the cervical cancer clinical target volume) is improved.

[0035] The method can realize automatic segmentation of the cervical cancer clinical target volume, greatly reduces the time and workload of manual delineation of doctors, improves the work efficiency, avoids segmentation differences caused by human factors, improves the consistency and accuracy of the segmentation results, and provides reliable technical support for radiotherapy of cervical cancer.

[0036] The application will be further described below in combination with the drawings and specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The flowchart of the cervical cancer clinical target volume image segmentation method of the application is shown in the figure.

[0038] In the figure, the flow from left to right is: obtaining a CT image of a patient; the data preprocessing link includes image normalization, cropping and scaling, labeling and data set division; then entering the Dense U-Net neural network model construction stage; then training the model, involving forward propagation, loss calculation, backward propagation parameter update and parameter adjustment using the verification set; finally, the test set image is input into the trained model for cervical cancer clinical target volume image segmentation.

[0039] Figure 2 Figure 1 is a schematic diagram of the deep learning network architecture based on Dense Block and Skip connection of the present application.

[0040] In the figure: (a) is the encoder and decoder part. The encoder contains multiple Dense Blocks and transition layers, and the convolutional layers in the Dense Block are in series dense connection state; the decoder part is symmetrical to the encoder; the middle part is the skip connection, which splices the feature maps of the corresponding layers of the encoder and the decoder. (b) is the Dense_Block part. DETAILED DESCRIPTION

[0041] The embodiment provides a cervical cancer clinical target volume image segmentation method based on a Dense U-Net neural network, to solve the problems of low segmentation accuracy and information loss of the cervical cancer clinical target volume in the prior art, and to realize more accurate and efficient segmentation of the cervical cancer clinical target volume.

[0042] As shown in Figure 1 : The flow from left to right is as follows: obtaining the CT image of the patient; the data preprocessing link includes image normalization, cropping and scaling, labeling and dataset division; then entering the Dense U-Net neural network model construction stage; then training the model, involving forward propagation, loss calculation, backward propagation parameter update and parameter adjustment using the validation set; finally, inputting the test set image into the trained model to segment the cervical cancer clinical target volume image. The feature transmission mode is to directly concatenate the features of all previous layers and transmit them to the next layer, rather than having an arrow pointing to all the following layers from each previous layer. A cervical cancer clinical target volume image segmentation method based on a Dense U-Net neural network includes the following steps:

[0043] 1. Data preprocessing: collect the CT images of cervical cancer patients, and perform normalization processing on the images to normalize the pixel values of the images to the range of [0, 1], and at the same time, perform cropping and scaling operations on the images to make them have a uniform size; manually label the clinical target volume to generate the corresponding labeled image, and divide the original image and the labeled image into a training set, a validation set and a test set. In practice: collect CT images of 100 cervical cancer patients, and normalize the pixel values of the images to the range of [0, 1] through the formula:

[0044]

[0045] wherein (x) is the original pixel value, x min and x maxrespectively. The image is cropped and scaled to the size of 256x256. The clinical target volume is manually labeled by an experienced deputy chief physician of tumor radiotherapy department, and a labeled image is generated. The 100 data are divided into a training set, a validation set and a test set according to a ratio of 7:1:2.

[0046] 2. Building a Dense U-Net neural network model: the Dense U-Net neural network model comprises an encoder part, a decoder part and a skip connection part; the encoder part is composed of multiple dense blocks and transition layers, each dense block is composed of multiple convolution layers, in the dense block, the input of each convolution layer is the splicing of the outputs of all previous convolution layers, in this way, the feature information is fully utilized and the information loss is reduced; the transition layer is used to reduce the resolution of the feature map and reduce the amount of calculation; the decoder part is symmetrical to the encoder part and also composed of multiple dense blocks and transition layers, for upsampling the feature map and restoring the resolution of the image; the skip connection part splices the feature maps of the corresponding layers in the encoder and the feature maps of the corresponding layers in the decoder, realizes the reuse of features and enhances the network's ability to capture detailed information.

[0047] In this embodiment: the encoder part contains 4 dense blocks, each dense block contains 4 convolution layers, the convolution kernel size is 3x3, the step is 1, and the padding is 1; a transition layer is arranged between each dense block, the transition layer contains a convolution layer, the convolution kernel size is 1x1, the step is 2, and is used to reduce the resolution of the feature map. The decoder part is symmetrically arranged with the encoder. The skip connection part splices the feature maps of the corresponding layers in the encoder and the feature maps of the corresponding layers in the decoder in the channel dimension. As shown in Figure 2 .

[0048] 3. Training the Dense U-Net neural network model: input the original image of the training set into the built Dense U-Net neural network model, calculate the output result of the model through forward propagation; according to the output result of the model and the labeled image, calculate the loss function, the present application adopts the combination of cross-entropy loss function and Dice coefficient loss function as the loss function, in order to better balance the class imbalance problem and improve the segmentation accuracy; update the parameters of the network through the back propagation algorithm, constantly adjust the weights and biases of the network, and minimize the loss function; in the training process, the model is verified by using the validation set, the training parameters such as learning rate and batch size are adjusted according to the result of the validation set, to prevent the model from overfitting, when the loss function on the validation set no longer decreases or reaches the preset training times, the training is stopped, and the trained Dense U-Net neural network model is obtained.

[0049] In specific, the original image of the training set is input into the model, and the output result is calculated by forward propagation. The loss function (L) adopts the weighted sum of cross-entropy loss function L ce and Dice coefficient loss function L dice , that is:

[0050] L = a · L ce + (1-a) · L dice

[0051] wherein a takes the value of 0.5. The network parameters are updated using the Adam optimizer, the initial learning rate is set to 0.001, and the batch size is 8. During the training process, the model is verified using the validation set every 5 epochs, and the learning rate and other parameters are adjusted according to the loss function and segmentation accuracy on the validation set. When the loss function on the validation set no longer decreases for 10 consecutive epochs, the training is stopped, and the trained model is obtained.

[0052] 4. Cervical cancer clinical target area image segmentation: the original image of the test set is input into the trained Dense U-Net neural network model, and the output result of the model is obtained by forward propagation. The output result is a segmentation image with the same size as the input image, and each pixel point in the segmentation image corresponds to a class label, which is used to represent whether the pixel point belongs to the cervical cancer clinical target area, thereby realizing the segmentation of the cervical cancer clinical target area image.

[0053] Through the above implementation, the cervical cancer clinical target area image segmentation based on the Dense U-Net neural network can be effectively realized, and the accuracy and efficiency of the segmentation are improved.

[0054] In this embodiment, in the DenseNet, all previous layers are connected as input, and its formula is x e = H e ([x0, x1,... x e-1 ])

[0055] wherein H(.) represents a nonlinear transformation function, which is a combination operation, which may include a series of BN, ReLU, Pooling and Conv operations.

[0056] Batch Normalization(BN) is a technique proposed by Google to accelerate training, which is basically a standard in many networks. It normalizes the data through four steps(mean, variance, gamma, beta) to facilitate the subsequent ReLU to effectively filter features. BN is usually placed before ReLU to ensure that ReLU can correctly process the adjusted data and avoid information loss or misjudgment. At the same time, the existence of BN makes the bias of the convolution layer optional, because the beta parameter of BN can replace the function of bias.

[0057] 2.Rectified Linear Unit(ReLU)

[0058] The effect of ReLU activation function is to set negative values to zero, which can be understood as eliminating information unrelated to features, and its output is useful information and a large area of 0. It is widely used in neural networks, which can effectively alleviate the gradient vanishing problem and speed up the training of the network.

[0059] 3.Pooling

[0060] Pooling layer is mainly used to reduce the dimension of feature map, reduce the amount of calculation, and enhance the robustness of the model. Common pooling operations include max pooling and average pooling. In Convlayers, the kernel_size of the pooling layer is usually 2, and the stride is 2, so that each MxN matrix after the pooling layer will be changed to (M / 2)*(N / 2) size.

[0061] 4.Convolution(Conv)

[0062] Convolution layer is the core component of Convolutional Neural Network(CNN), which extracts the features of data by sliding the convolution kernel on the input data. In multi-channel image convolution, the number of output image channels is always equal to the number of convolution kernels, regardless of the number of input image channels. Doing 1x1 convolution on multi-channel images is equivalent to connecting the original independent channels in the image together.

[0063] Relationship and combined application between technologies

[0064] 1.BN and Conv combined

[0065] Folding BatchNorm is not an operation that only exists in quantization. In general networks, BN can also be merged into Conv to speed up network inference. The merging process is to integrate the operations of BN into Conv to reduce computational overhead. Assuming that there is a trained Conv and BN, through certain formula transformation, the parameters of BN can be integrated into the parameters of Conv. When quantizing, the same method can be used, and the consistency of scale and zeropoint before and after ReLU is considered to maintain consistency before and after quantization. However, activation functions other than ReLU may not be suitable for this merging method.

[0066] 2. Combination of BN, ReLU, Pooling, and Conv in CNN

[0067] In CNN, a convolutional layer plus an activation layer (with BN) forms a basic unit. After several such operations, there will be a pooling operation, and then this process is repeated until the fully connected layer (usually one layer, but there may be multiple layers), and then output. The activation function ReLU can be replaced by other activation functions, and the order of activation and BN operations in the fully connected layer can be interchanged. The general output mapping uses the sigmoid function. In the entire Conv layers, the conv and relu layers do not change the input and output size, and only the pooling layer makes the output length and width become 1 / 2 of the input.

[0068] Application scenarios

[0069] These technologies are mainly applied in the field of image processing, such as in the Faster R-CNN model for target detection, which contains multiple conv layers, relu layers, and pooling layers in the Conv layers. Through the combination of these layers, the features of the image can be effectively extracted, and a feature map corresponding to the original image can be generated, thereby realizing the function of target detection.

[0070] "concat" is usually an abbreviation of "concatenate" in programming and data processing, which refers to the operation of connecting multiple string, array, list, and other data sequences into one. The following introduces the "concat" operation in different programming languages:

[0071] Python

[0072] In Python, the + operator or the join() method can be used to implement the concatenation operation for strings; for lists, the + operator or the extend() method can be used.

[0073] For example Figure 2As shown, (a) is the encoder and decoder part. The encoder contains multiple dense blocks and transition layers, and the convolutional layers in the dense blocks are in series dense connection state; the decoder part is symmetrical to the encoder; the middle part is a jump connection that splices the feature maps of the corresponding layers of the encoder and the decoder.

[0074] As shown in (b), this is a deep learning network architecture diagram based on dense connection (Dense Block) and jump connection (Skip connection), divided into two parts. Figure 2

[0075] The upper part is also the overall flow part:

[0076] It shows the encoder-decoder (Encoder-Decoder) structure, which is used for image tasks (input and output are images).

[0077] Encoder (encoder): The input image is convolved (Conv) into a dense block (features are gradually extracted densely), combined with maximum pooling (Max_Pooling, compressing features and increasing the receptive field) and jump connection (Skip connection, transmitting intermediate features to avoid information loss), gradually compressing the image size and extracting high-level features.

[0078] Decoder (decoder): Receive the output of the encoder, upsample (Up_Sampling, restore image size), dense block, jump connection, and finally output the processed image through convolution, dropout (regularization, prevent overfitting), to achieve the image task process of "compression-recovery / processing" (such as denoising, super-resolution, etc.).

[0079] The lower part mainly includes Dense Block and jump connection:

[0080] Focus on the features of Dense Block and jump connection:

[0081] In the dense block (Dense Block), the feature map (such as \(C_{input}^D\)) is convolved (Conv), batch normalized (BN), and activated ReLU, layer by layer, and densely spliced (feature reuse, each layer fuses all previous layer outputs), to generate subsequent feature maps (\(H_2^D\), \(H_1^D\), etc.).

[0082] ​Skip connection transfers features across layers (e.g. \(C_{input}^D\) is directly connected to the decoder side), allowing the decoder to directly access the fine-grained features from the early stages of the encoder, assisting in accurately restoring image details and solving the "feature loss" problem in encoding-decoding.

[0083] Overall, this is a typical dense network-based image inpainting / reconstruction model that strengthens feature reuse with dense connections and complements information transmission with skip connections, adapting to image tasks that require detail preservation.

Claims

1. A method for segmenting clinical target images of cervical cancer, characterized by: The following steps are involved: Step 1: Data preprocessing: Collect CT images of cervical cancer patients and normalize the images; Manually annotate the clinical target area, generate corresponding annotated images, and divide the original images and annotated images into training sets, validation sets, and test sets; Step 2: Build a Dense U-Net neural network model; Step 3: Train the Dense U-Net neural network model; Step 4: Segmentation of the cervical cancer clinical target image; input the original image into the trained Dense U-Net neural network model, and obtain the output result of the model through forward propagation. The output result is a segmented image with the same size as the input image. Each pixel in the segmented image corresponds to a category label, which is used to indicate whether the pixel belongs to the cervical cancer clinical target area, thereby realizing the segmentation of the cervical cancer clinical target image.

2. The method for segmenting clinical target images of cervical cancer according to claim 1, characterized in that: In step 1, when normalizing the image: the pixel values ​​of the image are normalized by the formula: Normalized to the range [0,1], where (x) is the original pixel value, x min and x max are the minimum and maximum pixel values ​​in the image, respectively.

3. The method for segmenting clinical target images of cervical cancer according to claim 2, characterized in that: In step 1, the image is normalized, cropped and scaled to a size of 256×256.

4. The method for segmenting clinical target images of cervical cancer according to claim 3, characterized in that: In step 1, an experienced oncologist in the radiation oncology department manually labels the clinical target area to generate a labeled image.

5. The method for segmenting clinical target images of cervical cancer according to claim 3, characterized in that: The step 1 further includes dividing the 100 data sets into a training set, a validation set, and a test set in a ratio of 7:1:

2.

6. The method for segmenting clinical target images of cervical cancer according to claim 1, characterized in that: In the step 2, The Dense U-Net neural network model includes an encoder part, a decoder part and a skip connection part; The encoder part includes at least one dense block and a transition layer for reducing the resolution of the feature map and reducing the amount of computation; each dense block is composed of at least two convolutional layers, and in the dense block, the input of each convolutional layer is the concatenation of the outputs of all previous convolutional layers; The decoder part is used to upsample the feature map and restore the resolution of the image; it is symmetrical with the encoder part and includes at least one dense block and a transition layer; The jump connection part concatenates the feature map of the corresponding layer in the encoder with the feature map of the corresponding layer in the decoder.

7. The method for segmenting clinical target images of cervical cancer according to claim 6, characterized in that: In step 2, the encoder part includes 4 dense blocks, each dense block includes 4 convolutional layers, the convolution kernel size is 3×3, the step size is 1, and the padding is 1; a transition layer is set between each dense block, the transition layer includes a convolution layer, the convolution kernel size is 1×1, and the step size is 2.

8. The method for segmenting clinical target images of cervical cancer according to claim 1, characterized in that: The step 3 includes: Step 301: Input the original images of the training set into the constructed Dense U-Net neural network model, and calculate the output of the model through forward propagation; Step 302: Calculate the loss function based on the output of the model and the labeled image; Step 303: Update the network parameters through the back propagation algorithm, and continuously adjust the network weights and biases to minimize the loss function; During the training process, the model is verified using the validation set, and the training parameters are adjusted according to the results of the validation set. When the loss function on the validation set no longer decreases or reaches the preset number of training times, the training is stopped to obtain the trained Dense U-Net neural network model.

9. The method for segmenting clinical target images of cervical cancer according to claim 8, characterized in that: During the training process of step 3, the loss function L adopts the cross entropy loss function L ce and Dice coefficient loss function L dice The Adam optimizer was used to update the network parameters, with the initial learning rate set to 0.001 and the batch size to 8. The model was validated using the validation set after every 5 epochs, and the parameters, including the learning rate, were adjusted based on the loss function and segmentation accuracy on the validation set. Training was stopped when the loss function on the validation set stopped decreasing for 10 consecutive epochs.

Images