A Deep Learning-Based Method for Breast Calcification Analysis

Abstract

This invention relates to a deep learning-based method for breast calcification analysis. By enhancing image representations, the method improves the model's ability to identify blurred and microcalcifications. An improved image segmentation network, UNeXt, based on a convolutional multilayer perceptron, effectively segments breast calcifications from images, significantly reducing computational overhead and costs. Combining deep learning with breast calcification features for analysis assists doctors in image interpretation and reduces errors. For the breast calcification analysis model, Focal loss + Dice loss are used as initial training functions, followed by HD loss for further training, thereby improving the model's learning ability. The deep learning-based breast calcification analysis method disclosed in this invention can serve as an auxiliary screening tool for early breast cancer, assisting doctors in better identifying and defining the morphology, size, and distribution of breast calcifications within the medical system.

Description

Technical Field

[0001] This invention belongs to the field of computer-assisted medical screening, specifically relating to a deep learning-based method for analyzing breast calcification. Background Technology

[0002] According to the latest statistics from around the world, breast cancer is considered one of the deadliest cancers. It is one of the most common cancers among women and the second leading cause of cancer death in women. Breast cancer is a malignant tumor caused by uncontrolled cell division. Currently, one of the medical prevention and treatment methods for breast cancer is regular screening for early detection and intervention.

[0003] Mammography refers to an X-ray examination of the breast. Mammography is one of the main methods for early detection and screening of breast cancer.

[0004] While breast cancer lesions come in many different forms, breast calcifications, as one of the four most common breast lesions, are often not to be ignored. Breast calcifications are calcium deposits that form in the breast tissue of women and are a sign of early-stage breast cancer, usually associated with ductal carcinoma in situ. However, calcifications can also occur in invasive cancers and are defined as a major sign of malignancy in breast imaging therapy. Approximately 41.2% of women in screening programs are described as having calcifications as the sole indication of cancer.

[0005] Assessing breast calcifications helps determine patient management. The morphology, size, and distribution of calcifications are important indicators to determine whether they are benign, nonspecific, and whether further imaging or biopsy is needed. Larger calcium deposits in breast tissue are called macrocalcifications; these particles are larger than 2 mm and are usually benign lesions. Calcifications smaller than 0.5 mm are called microcalcifications. Microcalcifications are also considered tiny calcium deposits and appear as white spots on mammograms. Microcalcifications are not always a sign of malignancy, but their presence can be a sign of breast cancer.

[0006] Because the sensitivity of calcification detection during initial mammogram screening is low, computer-aided diagnostic methods can provide clinicians with better decision support for early detection and analysis of calcifications. Deep learning-based computer-aided diagnostic algorithms have proven to provide an effective and robust solution for the automated analysis of breast cancer in mammograms. Image segmentation techniques are required to detect breast calcifications. Among deep learning-based image segmentation techniques, UNet is considered an efficient and robust segmentation method. UNet was a landmark work over the past decade, demonstrating the efficiency of encoder-decoder convolutional networks with skip connections in medical image segmentation. In recent years, UNet has become the backbone of almost all leading medical image segmentation techniques. UNet++, V-Net, Y-Net, and their latest extensions such as TransUNet are leading medical image segmentation methods in recent years.

[0007] The presence of numerous network parameters, complex calculations, and slow processing speed, coupled with the fact that mammogram images are typically large while calcifications are relatively small, presents challenges for the detection and analysis of microcalcifications. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention aims to provide a deep learning-based method for breast calcification analysis. Based on an improved image segmentation network, UNeXt, using convolutional multilayer perceptrons, it effectively segments breast calcifications from images, significantly reducing computational overhead and costs. Combining deep learning with breast calcification features for analysis assists doctors in image interpretation and reduces errors. This method can serve as an auxiliary screening tool for early breast cancer, helping doctors better identify and define the morphology, size, and distribution of breast calcifications within the medical system.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] A deep learning-based method for analyzing breast calcifications, the method comprising the following steps:

[0011] S11. Obtain the X-ray image data of the breast to be analyzed, process it into a preset format and size image that matches the format of the breast calcification analysis model;

[0012] S12. Perform data enhancement on the mammogram data to be analyzed after format processing;

[0013] S13. Input the data-enhanced breast X-ray image data to be analyzed into the breast calcification analysis model, and use the breast calcification analysis model to perform inference to obtain the image segmentation result.

[0014] S14. The image segmentation results are spliced ​​according to the cropping positions. After removing the preset edge pixels at the overlapping positions, the images are directly added together to obtain the final image segmentation result. Breast calcification analysis is performed based on the final segmentation result. The breast calcification analysis includes counting the number of calcifications, the number of microcalcifications, the number of macrocalcifications, the calcification density, the density around the calcifications, and the calcification clusters.

[0015] Furthermore, the format processing described in step S11 includes converting the mammogram data to be analyzed into a 16-bit PNG image, cropping the entire image to a size of 512*512, and setting the overlapping area to a first preset value.

[0016] Furthermore, the data augmentation method in step S12 includes at least one of image scaling, image sharpening, image grayscale value transformation, image blurring, image aliasing, and image copy and paste strategies.

[0017] Furthermore, step S13 includes averaging the multiple results to obtain the image segmentation result when multiple results are inferred from an enhanced image.

[0018] Furthermore, the breast calcification analysis model was established using the following method:

[0019] S1. Collect historical mammogram data containing calcified lesions, and annotate the collected mammogram data with calcifications to form corresponding label data;

[0020] S2. Preprocess the label data, the preprocessing including cropping the mammogram data and its corresponding label data and performing data augmentation on the cropped data;

[0021] S3. Establish a breast calcification analysis model based on the data augmented, and train the breast calcification analysis model using the improved UNeXt algorithm.

[0022] Furthermore, step S2 includes the following sub-steps:

[0023] S21: Binarize the label data, and use a matrix of all 1s of a preset dilation operator size to calculate the calcification point positions of the label data to obtain the dilated label data;

[0024] S22. Based on the location of the target connected component of the expanded label data, the mammogram data and its corresponding label data are cropped using a size of 512*512, and the overlapping area is set to the first preset value.

[0025] S23. The cropped mammogram data and its corresponding cropped label data are randomly divided into N parts, of which m parts are used as the training set of the mammary calcification analysis model and n parts are used as the test set of the mammary calcification analysis model, N = m + n.

[0026] S24. Perform data augmentation on the training set and the test set respectively.

[0027] Furthermore, step S3 includes the following sub-steps:

[0028] S31. Set the parameters of the breast calcification analysis model. The setting of the breast calcification analysis model parameters includes setting the image size to 512*512, setting the channels to channels: [64, 64, 128, 256, 512, 1024, 2048], and setting the loss function to Dice Loss+Focal Loss and HD Loss.

[0029] S32. Divide the training set data into N parts;

[0030] S33. Train the breast calcification analysis model using N-fold cross-validation, with m samples used for training and n samples used for validation.

[0031] S34. Test the trained breast calcification analysis model. A total of N breast calcification analysis models are saved. Test each model on the data-enhanced test set. Then, average the multiple segmentation result images to obtain the final result. Evaluate the effect of the breast calcification analysis model based on this result.

[0032] Furthermore, in step S33, the Focal Loss+Dice Loss loss function is used for several training iterations, and the best model on the validation set is saved. While focusing on the calcified lesion target, the HD Loss function is used to correct the boundary problem of the calcified lesion. Then, the HD Loss loss function is used to train the saved model, and the best model on the validation set is saved.

[0033] Furthermore, the breast calcification analysis model uses an improved UNeXt medical image segmentation network for image segmentation. Each layer of the improved UNeXt medical image segmentation network uses only one convolution operation. The improved UNeXt medical image segmentation network inputs the image into the first convolutional layer, which has 64 channels and a stride of 1. The first feature map is not downsampled during the skip connection, and the input image size is fixed at 512*512. The number of channels in each layer of the encoding process is set to 64, 64, 128, 128, 256, 512, 1024, and 2048, respectively. The original convolutional kernels in the encoding process are replaced with fused-MBConv, a compression reward / penalty network module is added, and labeled MLP block labels and projective convolutional features are used. The input channels are shifted when inputting into the MLP.

[0034] Furthermore, the improved UNeXt medical image segmentation network structure consists of convolutional layers, fused-MBConv layers, and Tok MLP layers. The encoding structure of the improved UNeXt medical image segmentation network includes one convolutional layer, four fused-MBConv layers, and two Tok MLP structures; the decoding structure of the improved UNeXt medical image segmentation network includes two Tok MLPs and four convolutional layers.

[0035] The advantages of this invention are as follows: Employing a deep learning-based breast calcification analysis method disclosed herein, to effectively learn breast calcification characteristics, data augmentation techniques are employed based on these characteristics, including image scaling, image grayscale transformation, image sharpening, image blurring, image aliasing, and image copy-paste strategies. Image scaling helps the breast calcification analysis model adapt to small-sized calcifications; image blurring, image aliasing, and grayscale transformation effectively improve the detection of calcifications accompanying masses and high-density glandular calcifications; the image copy-paste strategy involves copying calcification points from an image, counting the number of connected components, selecting connected components at a random ratio (0-1), and then pasting them into other images. This method not only significantly increases the number of labeled calcification points but also effectively improves the calcification segmentation efficiency of the breast calcification analysis model due to the inherent similarity of calcifications. This enables radiologists to effectively assess the morphology, size, and distribution of calcifications.

[0036] For the breast calcification analysis model, Focal loss + Dice loss are used as the loss function for initial training, and then HD loss is used to continue training the model to improve its learning ability. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating a method for establishing a breast calcification analysis model based on deep learning, as shown in Embodiment 1 of the present invention.

[0038] Figure 2 This is an illustration of the effect of effectively annotating calcifications in breast X-ray images containing calcified lesions, as shown in Embodiment 1 of the present invention.

[0039] Figure 3 This is a schematic diagram of the improved UNeXt medical image segmentation network structure shown in Embodiment 1 of the present invention;

[0040] Figure 4 This is a flowchart illustrating a deep learning-based breast calcification analysis method according to Embodiment 2 of the present invention.

[0041] Figure 5 This is a block diagram of the analysis results obtained by image segmentation according to the analysis method shown in Embodiment 2 of the present invention. Detailed Implementation

[0042] The present invention will now be further described with reference to the accompanying drawings and specific embodiments.

[0043] Example 1

[0044] like Figure 1 As shown in the figure, an embodiment of the present invention illustrates a method for establishing a breast calcification analysis model based on deep learning. The method includes the following steps:

[0045] Step S1: First, collect historical mammogram data containing calcified lesions, and effectively annotate the collected mammogram data with calcifications to form corresponding label data.

[0046] The collected historical mammogram data consisted of mammograms with breast calcification lesions. In this embodiment, the open-source software used for annotation was 3D Slicer. Since most calcifications are small-area calcifications, the level tracing function in the software was used as much as possible to ensure the accuracy of the annotation edges, thereby reducing human annotation errors and lowering the difficulty and error of learning the breast calcification analysis model.

[0047] In this embodiment, a total of 892 cases were collected, each containing more than four images, all with MLO (internal and external oblique views) and CC (axial or cephalottic views). A total of 1968 images with calcified lesions were collected, with an image size of 2832*3200. The original data was saved as 12-bit images. In this embodiment, the data was expanded to 16-bit PNG images for annotation. The specific annotation effect is as follows. Figure 2 As shown.

[0048] Step S2: Preprocess the label data. Step S2 involves preprocessing the label data, including cropping the mammogram image data and its corresponding original label data, and enhancing the cropped data. Step S2 includes the following sub-steps:

[0049] Step S21: Binarize the label data, and then use a matrix of all 1s of the preset dilation operator size to calculate the calcification point positions of the label data to obtain the dilated label data.

[0050] Since breast calcifications typically occupy only about 1 / 1000 of the entire image area, the data is extremely unbalanced, and the calcifications are scattered, making it almost unnecessary to consider global features. Therefore, in this step, a 128*128 matrix of all-ones is used to dilate the calcification points in the label data.

[0051] Step S22: Based on the location of the target connected component in the expanded label data, the mammogram data and its corresponding original label data are cropped using a preset cropping size.

[0052] Calculate the connected components with a value of 1 in the dilated label data, and then crop the mammogram images and their corresponding original label images according to the location of the connected components. The cropping size is 512*512, and the overlap area is set to 100. The result ensures that calcification exists in each cropped image, and the calcification is not at the edge of the cropped image.

[0053] Step S23: The cropped mammogram data and the corresponding cropped label data are randomly divided into 5 parts, of which 4 parts are used as the training set for the mammary calcification analysis model and 1 part is used as the test set for the mammary calcification analysis model.

[0054] Continuing from the previous example, after cropping, the number of images obtained is 9726 and their corresponding labels. The data is randomly divided into a training set and a test set with a ratio of 4:1, corresponding to 7781 and 1945 images respectively.

[0055] Step S24: Perform data augmentation on the training set and the test set respectively.

[0056] To fully utilize the limited labeled data, data augmentation is performed on both the training and test sets. Data augmentation methods include: image scaling, image grayscale transformation, image sharpening, image blurring, image aliasing, and image copy-paste strategies. Data augmentation step S24 includes the following sub-steps:

[0057] Step S241: Image scaling, with scaling factors of 0.2, 0.5, 0.8, 2, 3 and 5 respectively. For example, if the data size is 2832*3200, scaling by 2 will result in a scaled structure of 4005*4525, and scaling by 0.2 will result in a scaled structure of 1266*1431.

[0058] Step S242: Image sharpening. The Robert operator, Sobel operator, and Scharr operator are used to calculate the image, which can effectively improve the boundary detection effect.

[0059] Step S243: Image grayscale value transformation is based on histogram transformation. First, the image histogram is obtained, and then the image is transformed by stretching or compressing the image frequency domain. The methods used include histogram equalization; histogram stretching or compression: randomly increasing or decreasing the grayscale value range of the bright part, and decreasing or increasing the grayscale value range of the corresponding other parts; histogram compression can improve the model's ability to recognize fuzzy calcification.

[0060] Step S244: Image blurring is mainly achieved by adding noise and filtering. The noise includes Gaussian noise, speckle noise and Poisson noise, and the noise setting parameters are randomly generated. The filtering includes median filtering, Vienna filtering, mean filtering and wavelet filtering.

[0061] Step S245: Image aliasing, including mixup and cutout. Mixup mixes two random samples proportionally and adds the segmentation results together. Cutout occludes the image using a rectangle of random size. Within the rectangle, all values ​​are set to 0, and if there is a corresponding label, it is also set to 0.

[0062] Step S246: Image copy and paste strategy. Copy the calcification points in the image, then count the number of connected components, select a certain number of connected components using a random ratio (0~1), and then copy the calcification points where the selected connected components are located from the breast image and paste them into other breast images. This method not only greatly increases the number of labeled calcification points, but also effectively improves the model's performance due to the similarity of the calcifications themselves.

[0063] Step S3: Establish a breast calcification analysis model based on the enhanced training and test sets, and train the model using the improved UNeXt algorithm. Step S3 includes the following sub-steps:

[0064] Step S3.1: Set the model parameters. In this embodiment, the image size is set to 512*512, the number of training iterations is set to 400+100, the cross-validation folds are set to 5, the batch size is set to 4, the optimizer is set to Adam, the learning rate is set to le-4, the momentum is set to 0.9, and the channels are set to [64, 64, 128, 256, 512, 1024, 2048]. The loss function is set to Dice Loss + Focal Loss and HD Loss. The training machine parameters are: Intelli9-11900K, RTX A4000 16G VRAM, 32G RAM.

[0065] Step S3.2: Divide the data into five parts.

[0066] Step S3.3: Train the model using five-fold cross-validation, with four copies used for training and one copy for validation each time. First, use Focal Loss + Dice Loss for 400 training iterations. Save the best model on the validation set. Then, while focusing on the calcified lesion target, use HD Loss to correct the calcified lesion boundary. Next, train the saved model using HD Loss for 100 iterations, keeping other parameters unchanged. Save the best model on the validation set as well.

[0067] Step S3.4: Test the trained breast calcification analysis model. A total of five breast calcification analysis models are saved. Each model is tested on the data-enhanced test set. The resulting cropped images are then averaged to obtain the final result. Finally, the model performance is evaluated.

[0068] like Figure 3 As shown, for calcification features, the breast calcification analysis model disclosed in this embodiment of the invention uses an improved UNeXt medical image segmentation network for image segmentation. The network structure of the improved UNeXt medical image segmentation network consists of convolutional layers, fused-MBConv layers, and Tok MLP layers. The encoding structure of the improved UNeXt medical image segmentation network includes one convolutional layer, four fused-MBConv layers, and two Tok MLP structures; the decoding structure of the improved UNeXt medical image segmentation network includes two Tok MLPs and four convolutional layers.

[0069] The following improvements were made to the UNeXt medical image segmentation network:

[0070] To improve segmentation speed, compared to the Unet network, the improved UNeXt network uses only one convolution operation per layer. In order to extract features of small calcifications, the image is first input into the first convolutional layer with 64 channels and a stride of 1. The first feature map is not downsampled when performing skip connections, thereby maximizing the preservation of small calcification feature maps.

[0071] To extract more features, the input image size was fixed at 512*512. The number of channels in each layer of the encoding process were 64, 64, 128, 256, 512, 1024, and 2048, respectively, and the decoding process was similar. Figure 3 As shown.

[0072] To learn image features more accurately, this invention replaces the convolutional kernels in the encoding process with fused-MBConv. Compared to the traditional Conv, it adds a compression reward and punishment network module, Squeeze-and-Excitation Blocks (SE), which can learn image features more accurately.

[0073] To improve model inference speed, this invention utilizes Unext, an image segmentation network based on an MLP (Multilayer Perceptron), specifically the Fused-Conv+MLP structure, based on the UNeXt network. Tokenized MLP blocks (Tok MLP) are used to label and project convolutional features. To enhance performance, the input channels are shifted during MLP input, allowing for better learning of local feature distributions. This improved UNeXt network achieves better segmentation performance and significantly increased speed.

[0074] Example 2

[0075] like Figure 4 As shown in the figure, this invention illustrates a deep learning-based method for breast calcification analysis. The method is based on a deep learning-based breast calcification analysis model establishment method disclosed in Embodiment 1. The deep learning-based breast calcification analysis method includes the following steps:

[0076] Step S11: Obtain the X-ray image data of the breast to be analyzed, process it into a preset format and size image that matches the format of the breast calcification analysis model.

[0077] In this embodiment, the mammogram data to be analyzed is converted into a 16-bit PNG image, and the entire image is cropped to a size of 512*512, with the overlap area set to 100.

[0078] Step S12: Perform data augmentation on the mammogram data to be analyzed after format processing. The data augmentation process is the same as that during model training.

[0079] Step S13: Input the data-enhanced mammogram data to be analyzed into the mammary calcification analysis model, and use the mammary calcification analysis model for inference. Since data enhancement and five-fold cross-validation are used, a 512*512 image will produce multiple results. Average the multiple results to obtain the image segmentation result.

[0080] Step S14: Segmentation results were obtained for each 512*512 image. The segmentation results were then stitched together according to the cropping positions, with 20 pixels removed from the overlapping positions. The results were then added together to obtain the final image segmentation result.

[0081] like Figure 5 As shown in the figure, the number, area, and density of calcifications were statistically analyzed based on the segmentation results, resulting in the analysis results shown in the figure.

[0082] In summary, this invention discloses a deep learning-based method for breast calcification analysis. By enhancing image data, it improves the model's ability to identify blurred and microcalcifications. The improved network structure enhances image segmentation speed. Adding a 64-channel convolutional layer with a stride of 1 at the front end of the network maximizes the preservation of microcalcification feature maps, thus improving calcification detection accuracy. This allows for accurate calcification segmentation, enabling radiologists to statistically analyze calcification data, including the number of calcifications, microcalcifications, large calcifications, calcification density, calcification periphery density, and calcification clusters. This facilitates more accurate diagnosis and is significant for patient prognosis and follow-up. By segmenting breast calcifications, doctors can effectively analyze their morphology.

[0083] The methods and systems described in this invention are not limited to the embodiments described in the specific implementation. Other implementation methods derived by those skilled in the art based on the technical solutions of this invention also fall within the scope of technical innovation of this invention.

Claims

1. A deep learning-based breast calcification analysis method, characterized by, The method comprises the steps of: S11, acquiring breast X-ray image data to be analyzed, performing format processing thereon, and processing into a picture of a preset format and a preset size matching a breast calcification analysis model format; S12, performing data enhancement on the breast X-ray image data to be analyzed after format processing; S13, inputting the breast X-ray image data to be analyzed after data enhancement into the breast calcification analysis model, using the breast calcification analysis model for inference to obtain an image segmentation result; S14, splicing the image segmentation result according to a cropped position, wherein the overlapping position is directly added after removing a preset edge pixel to obtain a final image segmentation result, and breast calcification analysis is performed according to the final segmentation result, the breast calcification analysis including counting a calcification number, a micro-calcification number, a macro-calcification number, a calcification density, a calcification peripheral density, and a calcification cluster, The format processing in step S11 includes converting the breast X-ray image data to be analyzed into a 16-bit png image, cropping the whole image into a picture with a size of 512*512, and setting an overlapping area to a first preset value, The data enhancement method in step S12 includes at least one of image scaling, image sharpening, image gray value transformation, image blurring, image aliasing, and image copy-paste strategy, Step S13 includes averaging multiple results to obtain an image segmentation result when multiple results are inferred from one image after enhancement, The breast calcification analysis model uses an improved UNeXt medical image segmentation network for image segmentation, each layer of the improved UNeXt medical image segmentation network uses only one convolution operation, the improved UNeXt medical image segmentation network inputs an image into a first convolution layer, the first convolution layer has a channel number of 64 and a step of 1, a first feature map does not undergo down-sampling when performing a skip layer connection, and the input image size is fixed to 512*512, wherein the channel numbers of the layers in the encoding process are set to 64, 64, 128, 256, 512, 1024, and 2048, respectively, the convolution kernel in the original encoding process is replaced by a fused-MBConv, a marked MLP block is used to mark and project convolution features, and the channels of the input are translated when inputting the MLP. 2.The deep learning-based breast calcification analysis method of claim 1, wherein, The breast calcification analysis model is established by the following method: S1, collecting historical breast X-ray image data containing calcification lesions, performing calcification labeling on the collected breast X-ray image data, and forming corresponding label data; S2, pre-processing the label data, the pre-processing including cropping the breast X-ray image data and the corresponding label data, and performing data enhancement on the cropped data; S3, establishing a breast calcification analysis model according to the data after data enhancement, and using an improved UNeXt algorithm to train the breast calcification analysis model. 3.The deep learning based breast calcification analysis method of claim 2, wherein, Step S2 includes the following sub-steps: S21: binarize the label data, use a full 1 matrix of a preset dilation operator size to operate on the calcification point position of the label data to obtain dilated label data; S22, according to the position of the target connected domain of the dilated label data, the breast X-ray image data and the corresponding label data are cut to 512*512 size, and the overlapping area is set to a first preset value; S23, the cut breast X-ray image data and the corresponding cut label data are randomly divided into N parts, m parts are used as the training set of the breast calcification analysis model, n parts are used as the test set of the breast calcification analysis model, and N = m + n; S24, the training set and the test set are respectively subjected to data enhancement. 4.The breast calcification analysis method based on deep learning according to claim 3, characterized in that, Step S3 includes the following sub-steps: S31, setting the breast calcification analysis model parameters, including setting the image size to 512*512, setting the channel to channels: [64, 64, 128, 256, 512, 1024, 2048], and setting the loss function to Dice Loss+Focal Loss and HD Loss; S32, data division, the training set data is divided into N parts; S33, breast calcification analysis model training, N-fold cross-validation is used for breast calcification analysis model training, m parts are used for training, and n parts are used for verification; S34, test the trained breast calcification analysis model, save N breast calcification analysis models in total, test the data enhanced test set respectively, then average the obtained multiple segmentation result images to obtain the final result, and evaluate the breast calcification analysis model effect according to the final result.

5. The deep learning-based breast calcification analysis method of claim 4, wherein: In step S33, Focal Loss+Dice Loss loss function is used for several times of training, the best model on the validation set is saved, the focal calcification lesion target is corrected at the same time, and HD Loss function is used to correct the calcification lesion boundary problem, then the saved model is trained using HD Loss loss function, and the best model on the validation set is saved.

6. The deep learning-based breast calcification analysis method of claim 1, wherein: The network structure of the improved UNeXt medical image segmentation network is composed of convolutional layers, fused-MBConv layers and Tok MLP layers, the encoding structure of the improved UNeXt medical image segmentation network includes one convolutional layer, four fused-MBConv layers and two Tok MLP structures; and the decoding structure of the improved UNeXt medical image segmentation network includes two Tok MLPs and four convolutional layers.

Images