MRI (Magnetic Resonance Imaging) head and neck tumor segmentation method and system based on SCDU-Net model

Abstract

The invention discloses an MRI (Magnetic Resonance Imaging) head and neck tumor segmentation method and system based on an SCDU-Net model, an encoder module has multiple stages of down-sampling, each stage of down-sampling comprises a residual error convolution block and a coordinate attention module connected with the residual error convolution block, and the encoder module is used for extracting multi-level features of an image and reducing spatial resolution step by step; the bottleneck module is connected with the last-stage downsampling layer of the encoder module, comprises a dense cavity space pyramid pooling module and is used for extracting multi-scale context information of a target area through a plurality of cavity convolution branches with different expansion rates; and the decoder module is connected with the bottleneck module and receives the jump connection characteristics from the downsampling layers at all levels of the encoder module, and the decoder module has multi-level upsampling and is used for recovering the spatial resolution of the characteristic pattern step by step and outputting a segmentation result image with the same size as the input image. According to the method, the head and neck tumor MRI segmentation precision and robustness are improved.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, and relates to an MRI head and neck tumor segmentation method and system based on the SCDU-Net model. Background Technology

[0002] Head and neck tumors are malignant lesions formed by the abnormal proliferation of cells in the head and neck tissues. Because they are often diagnosed at a late stage and progress rapidly, treatment is difficult and the prognosis is poor. Early and accurate tumor image segmentation can significantly improve diagnostic accuracy and treatment effectiveness, thereby improving patient prognosis. Magnetic resonance imaging (MRI), as an important non-invasive imaging technique, is widely used in the diagnosis of tumors in the brain, nasopharynx, and oropharynx due to its high resolution and good soft tissue contrast. However, the traditional method of radiologists manually delineating lesions has many limitations, including long processing times, poor repeatability, and large inter- and intra-observer variability, affecting the reliability and consistency of segmentation results. Traditional segmentation methods, based on thresholding or region growing, are often sensitive to image noise and struggle to cope with the complex anatomical structures and heterogeneity of head and neck tumors, resulting in limited segmentation accuracy. In contrast, deep learning segmentation technology, with its powerful feature extraction capabilities and end-to-end automated processing advantages, can effectively overcome the shortcomings of traditional methods, significantly improving segmentation accuracy and efficiency, and providing reliable support for accurate diagnosis, treatment planning, and radiotherapy target delineation.

[0003] U-Net, through its symmetrical contraction path (encoder) and expansion path (decoder) combined with skip connections, has demonstrated outstanding performance in medical image segmentation. It effectively extracts multi-scale features while preserving boundary details, significantly improving the segmentation accuracy of structures such as brain tumors and livers. Its outstanding performance on medical datasets lays an important foundation for automated segmentation. However, U-Net still has some limitations: its decoding stage lacks an effective mechanism for allocating feature weights, making it difficult to highlight key features of the tumor region, resulting in insufficient differentiation between tumors and background. Especially with complex backgrounds, the model is prone to missing detailed tumor features; its localization ability is limited in segmenting small tumors and fine structures, easily leading to missed detections or missegments; the feature extraction of the bottleneck layer relies on a single scale, limiting the model's perception of multi-scale edge details. Head and neck tumors vary greatly in shape and size; without the fusion of features from different scales, the model may not be able to simultaneously consider large-scale backgrounds and local details, affecting segmentation accuracy. These problems are particularly evident in head and neck tumor segmentation tasks, impacting the model's clinical application effectiveness.

[0004] Therefore, it is necessary to design a new method and system for segmenting head and neck tumors to further improve the accuracy and robustness of MRI segmentation of head and neck tumors. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a method and system for MRI head and neck tumor segmentation based on the SCDU-Net model, solving technical problems such as insufficient feature extraction, difficulty in segmenting small targets, and inadequate information utilization in head and neck tumor MRI image segmentation. Specifically, this invention aims to improve the detection rate and boundary characterization ability of head and neck tumors (including small and irregularly shaped tumors) while maintaining the model's segmentation accuracy, and to enhance the model's adaptability to tumors of different sizes and its robustness to imbalanced data.

[0006] To achieve the above objectives, the present invention employs the following technical solutions: An MRI head and neck tumor segmentation system based on the SCDU-Net model includes an encoder module, a bottleneck module, and a decoder module. The encoder module has multi-level downsampling, each level of which includes a residual convolutional block and a coordinate attention module connected to the residual convolutional block, used to extract multi-level features of the image and progressively reduce the spatial resolution. The bottleneck module is connected to the last downsampling layer of the encoder module and includes a densely dilated spatial pyramid pooling module, used to extract multi-scale contextual information of the target region through multiple dilated convolutional branches with different dilation rates. The decoder module is connected to the bottleneck module and receives skip connection features from each level of downsampling layers of the encoder module. The decoder module has multi-level upsampling, used to progressively restore the spatial resolution of the feature map and output a segmentation result image with the same size as the input image.

[0007] As a preferred embodiment, the residual convolutional block includes at least two convolutional layers and a residual connection that adds the input of the residual convolutional block to the output of the convolutional layers, thereby enhancing feature extraction capabilities without increasing network depth.

[0008] As a preferred embodiment, the coordinate attention module extracts directional position information encoding by performing global pooling along the width and height directions of the input feature map, generating corresponding horizontal and vertical attention weight maps, and performing spatial attention enhancement on the input feature map; the output of the coordinate attention module is fused with the output of the one-channel attention SE module to simultaneously calibrate the feature map spatially and in channels.

[0009] As a preferred embodiment, the densely hollow spatial pyramid pooling module includes at least three hollow convolutional layers with different dilation rates. The input of each hollow convolutional layer includes the original input feature map of the densely hollow spatial pyramid pooling module and the output features of all previous hollow convolutional layers with dilation rates less than that of the hollow convolutional layer. By densely connecting layers one by one, the receptive field is expanded and multi-scale contextual information is obtained.

[0010] As a preferred embodiment, the encoder module also includes a convolutional coding branch and a Transformer coding branch in parallel, used to extract local features and long-range dependency features of the input image; the features extracted by the Transformer coding branch are fused with the features of the convolutional coding branch in the decoder module through an attention mechanism to enrich the global context information in the decoding stage.

[0011] As a preferred embodiment, at least two different levels of upsampling layers of the decoder module are respectively connected to a segmentation output submodule, which is used to output intermediate segmentation result images of corresponding resolution to provide multi-scale supervision signals during model training.

[0012] This invention also discloses a segmentation method for an MRI head and neck tumor segmentation system based on the SCDU-Net model, comprising the following steps: Step 1, Image Preprocessing: The original MRI image is intensity normalized and resampled to a predetermined uniform resolution to generate a single-channel two-dimensional image; random data augmentation processing is performed on the image, including random morphological transformation, random geometric transformation and random intensity transformation, and image slices that do not contain tumor regions are filtered out. Step 2, Feature Extraction: The preprocessed image is input into the encoder module of the segmentation system. Multi-level features are extracted through residual convolutional blocks and spatial information is enhanced by the coordinate attention module. After multi-level downsampling, a low-resolution deep feature map is obtained. Step 3, Context Feature Acquisition: Input the low-resolution feature map obtained in Step 2 into the densely dilated spatial pyramid pooling module of the bottleneck module, extract and fuse context features of different scales through dilated convolution branches with multiple dilation rates to obtain enhanced feature representations. Step 4, Upsampling Decoding: Input the feature representation obtained in Step 3 into the decoder module of the segmentation system, upsample step by step to restore spatial resolution, and fuse the skip connection features from the corresponding encoder layer to generate a segmentation result image with the same size as the original image; Step 5, Loss Calculation and Model Update: The segmentation result image is compared with the corresponding ground truth label at the pixel level. The combined loss value consisting of Dice loss and Focal loss is calculated, and the model parameters of the segmentation system are updated and adjusted based on the loss value through the backpropagation algorithm.

[0013] As a preferred approach, in step 4, the decoder module outputs intermediate segmentation result images of corresponding resolutions in multiple upsampling stages, calculates the loss value between each intermediate segmentation result image and the corresponding scaled real label, and finally weights and sums the loss values ​​of each resolution to obtain the total loss, thereby realizing multi-scale deep supervised training.

[0014] As a preferred approach, the calculation of the loss value further includes introducing a boundary loss term, which gives higher weight to the error at the edge of the target region in the segmentation result image.

[0015] As a preferred approach, in step 2, the image features are extracted in parallel by a convolutional encoder branch and a Transformer encoder branch, with the Transformer encoder branch used to extract long-range dependency features of the image; in step 4, the features extracted by the Transformer encoder branch and the features of the convolutional encoder branch are fused together through an attention mechanism and then decoded to obtain a more complete segmentation result image.

[0016] The present invention has the following advantages: First, residual convolutional blocks are introduced into each layer of the encoder to replace the traditional convolutional stacking, so that each encoding layer consists of two convolutional layers plus an identity skip connection. Through residual connections, the network depth is effectively increased, enhancing the feature extraction capability; on the other hand, the gradient vanishing problem caused by network deepening is alleviated, ensuring training stability and guaranteeing that key tumor feature information is preserved in each downsampling stage.

[0017] Second, by integrating an advanced attention module during feature extraction, the model's focus on tumor targets is enhanced. This invention combines the SE channel attention mechanism with the coordinate attention mechanism: on the one hand, the SE module is used to recalibrate channel features, dynamically adjusting the weights of each channel based on global information to highlight tumor-related features and suppress redundant information; on the other hand, an improved coordinate attention module is introduced to perform global pooling and encoding in both the horizontal and vertical directions of the feature map, introducing positional information encoding to highlight the spatial features of the tumor region. By fusing the CA and SE outputs to form a spatial-channel fusion attention, the model can simultaneously capture global dependencies and fine spatial locations, thereby significantly improving its ability to locate small lesions and its feature representation capabilities.

[0018] Third, a Dense Spatial Pyramid Pooling (DenseASPP) module is introduced at the bottleneck layer where the encoder and decoder meet. This module contains multiple parallel dilated convolutional branches, employing an increasing dilation rate to acquire feature representations under different receptive fields, and optimizing feature fusion through dense connections. In DenseASPP, each layer only receives the output of the previous layer and the original features with a smaller dilation rate than the current layer, thereby reducing redundancy and computational overhead, while maximizing the expansion of the receptive field to capture multi-scale contextual information. Through the DenseASPP module, the network can effectively perceive various scale changes in the tumor region, enrich contextual semantic information, and improve the robustness of segmentation for tumors of different sizes.

[0019] Fourth, a multi-scale supervision mechanism is employed during model training to apply auxiliary loss constraints to feature maps at different levels. Specifically, segmentation prediction branches are introduced at several stages of the decoder. The loss is calculated by comparing the prediction results at intermediate scales with the actual segmentation at the corresponding resolution, and then weighted and summed with the loss of the main output. This deep supervision strategy guides shallow feature learning, providing feedback for the segmentation task from an early stage, which helps the model better distinguish tumors from the background, especially improving the detection ability of small tumors and blurred boundaries.

[0020] V. To address the issues of imbalanced tumor and background categories and blurred tumor boundaries in medical segmentation, this invention designs a composite loss function that combines Dice loss and Focal loss. Focal loss adds a difficulty modulation factor to the standard cross-entropy loss, reducing the weight of easily classified samples and focusing on difficult samples, thereby mitigating the "submergence" of small tumors in the overall loss. Dice loss directly optimizes the overlap between predicted and actual segments (Dice coefficient), making it more sensitive to the overlap of tumor region boundaries. The linear weighted combination of these two loss functions allows the model to simultaneously achieve pixel-level accuracy and optimize region overlap, significantly improving segmentation performance under imbalanced data conditions. Attached Figure Description

[0021] Figure 1 This is a comparison chart of the labeling results using different methods. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings.

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0024] Example: An MRI head and neck tumor segmentation system based on the SCDU-Net model mainly includes the following functional modules: The MRI image acquisition and preprocessing module is used to acquire MRI image data of the head and neck region and perform necessary preprocessing on the raw images. Data acquisition can be from publicly available medical image databases or DICOM images obtained from clinical scans. In this embodiment, a set of head and neck tumor MRI datasets is used. Preprocessing includes: sampling 3D volume data layer by layer into two-dimensional slices; resampling and cropping the images, uniformly adjusting the resolution to 256×256 pixels; and intensity normalization to correct brightness differences between different scans. In addition, during the model training phase, the preprocessing module also performs data augmentation operations, including random morphological transformations, geometric transformations (rotation, scaling, flipping, etc.), and image intensity perturbations, thereby expanding the training samples and improving the model's generalization ability. For the training data, slices without tumors can be further filtered out to avoid interference.

[0025] SCDU-Net Deep Segmentation Module: This is the core of the system, used for tumor region segmentation in preprocessed MRI images. The SCDU-Net model employs an improved encoder-decoder architecture, comprising an encoder, bottleneck, and decoder. The encoder consists of four progressively downsampled convolutional layers, each containing two 3×3 convolutions, batch normalization, and activation units. Input features are directly passed to the layer output via residual connections, forming residual convolutional blocks. The introduction of residual blocks increases the effective network depth, but the presence of identity skip connections ensures smoother gradient propagation within each layer, preventing network degradation. Simultaneously, an attention submodule, including SE channel attention and coordinate attention, is fused after the convolution of each encoding layer to enhance the extraction of tumor region features. After a convolutional layer generates a feature map, the SE module first performs global channel compression and weight recalibration, calculating the contribution weight of each channel to the current task. Then, the coordinate attention module extracts the global spatial information of the feature map in the vertical and horizontal directions, generating a spatial attention map. Finally, the results of channel attention and spatial attention are multiplied and applied to the original feature map, forming a feature representation with both spatial and channel enhancements. This fusion attention mechanism allows the features extracted by the encoder to retain both global context and highlight key local parts, providing a more discriminative feature map for subsequent decoding.

[0026] After four layers of downsampling by the encoder, the feature map size is reduced from the original 256×256 to 16×16, while the number of channels expands from the initial smaller dimension to a higher dimension, accumulating rich high-level semantic information. Next, the bottleneck layer is entered, where this invention introduces the DenseASPP module to further extract multi-scale semantics. The bottleneck's DenseASPP consists of a series of parallel dilated convolutions (e.g., one 3×3 convolutional layer with dilation rates d=1, 3, and 5 each), and a dense connection strategy is used to optimize feature combination. Specifically, the first dilated convolution directly acts on the bottleneck input features, the second convolution accepts the bottleneck input and the connection of the first layer's output as input, the third layer accepts the bottleneck input and the outputs of the first two layers, and so on. The output of each layer is passed to the next layer after non-linear activation. High dilation rate convolutions can capture a larger range of global background, while low dilation rate convolutions preserve local details; dense connections ensure the gradual fusion of features at different scales, avoiding the redundancy and overhead of stitching together features at all scales at once. The output of the DenseASPP module contains tumor region features at multiple scales, and sufficient contextual information has been gathered on a 16×16 feature map, laying the foundation for accurate segmentation.

[0027] The decoder structure of SCDU-Net is symmetrical to the encoder, including four progressively upsampling layers. Each decoder layer first upsamples the output of the previous layer (e.g., by a factor of 2, gradually restoring the spatial size from 16×16 to the original image size), and then fuses it with the feature skip connection of the corresponding encoder layer (e.g., concatenating the output of the i-th layer of the encoder with the upsampled features of the i-th layer of the decoder along the channel dimension). The fused features are then subjected to several 3×3 convolutions and non-linear activations, gradually restoring spatial details while compressing the channel dimension, reconstructing a high-resolution segmentation feature map. Notably, in this invention, the attention enhancement features of each layer of the encoder are directly fed to the decoder through skip connections, allowing the decoder to refer to the key tumor information extracted during the encoding stage when reconstructing the image, thereby improving the restoration accuracy of small tumors and subtle boundaries. At certain layer outputs of the decoder, auxiliary output branches can also be set for multi-scale deep supervised training. These intermediate outputs are used to calculate the loss with the original labels (after appropriate downsampling), assisting the main output in optimization, so that the features at each level of the network are subject to supervised constraints, improving training efficiency and segmentation results.

[0028] The final decoder output undergoes a 1×1 convolution mapping to obtain a tumor probability map of the same size as the input MRI image. Pixel values ​​represent the confidence that a location belongs to a tumor. Thresholding the probability map (e.g., 0.5) typically yields a binary tumor region segmentation result. For multi-class cases (such as distinguishing between primary tumors and metastatic lymph nodes), the output layer can be configured with multiple channels to predict the probability of each class. Finally, the segmented region for each tumor class is obtained by selecting the highest probability or using multiple thresholds.

[0029] The results output and post-processing module processes the segmentation results and outputs them in a human-readable format. On one hand, it overlays the obtained 2D segmentation mask onto the original MRI image to generate a color-annotated map for doctors to view intuitively. On the other hand, when needed, it can perform 3D reconstruction of the 2D slices to restore the shape of the tumor in 3D space, for use in preoperative planning or radiotherapy dose design. Post-processing may also include morphological operations to smooth segmentation edges and fill small holes, improving the coherence and practical value of the segmented area.

[0030] The segmentation method of the present invention has the following steps: Step S1: Medical Image Acquisition. Acquire the head and neck MRI image data to be segmented. The data can come from a medical image database, a hospital image archive system, or real-time acquired patient MRI scans. Since MRI is typically three-dimensional volumetric data, this step may include extracting individual axial slice sequences from the volumetric data for subsequent processing.

[0031] Step S2: Image Preprocessing. Preprocessing operations are performed on the raw MRI images, including: normalizing image pixel intensity to a standard range to reduce brightness differences between different scans; performing quality optimization such as denoising and artifact correction as needed; and resampling or cropping the images to a uniform size (e.g., 256×256 pixels) to match network input requirements. If the original dataset contains multimodal images, a single modality (e.g., T2-weighted images) or multi-channel input (multimodal merging) can be selected for segmentation. During the training phase, this step also includes performing data augmentation, such as random rotation, flipping, scaling, random changes in brightness and contrast, and adding affine transformations or noise to generate more diverse training samples, thereby improving the model's adaptability to unknown data. After preprocessing, the image data and the corresponding manually annotated tumor mask (if used for training) are input into the next step.

[0032] Step S3: Model Training. With labeled data available, divide the dataset into training and validation sets according to a certain ratio (e.g., 80% training, 20% validation). K-fold cross-validation may be used if necessary to fully utilize the data. Train and optimize the SCDU-Net model using the preprocessed training set. During training, use the combined loss function (Dice + Focal) mentioned above, supplemented by multi-scale deep supervision (if multiple auxiliary output branches are enabled, the total loss is the weighted sum of the main output loss and each auxiliary loss). Optimization algorithms can include stochastic gradient descent (SGD) or adaptive moment estimation (Adam), iteratively updating the model parameters until the validation set loss converges or the accuracy metric reaches the expected level. To improve the model's generalization performance, regularization techniques such as early stopping and weight decay can be added during training. When using cross-validation, multiple model parameter files obtained from training can be obtained.

[0033] Step S4: Model Inference and Segmentation. The preprocessed MRI image to be tested is input into the trained SCDU-Net model to perform forward inference calculations, outputting the predicted tumor region. For multiple sub-models trained using cross-validation, an ensemble inference strategy can be used on the same input: the tumor probability map output independently by each model is obtained and averaged to obtain a final probability map that integrates insights from multiple models. This ensemble method can reduce the uncertainty of single-model predictions to some extent, resulting in more stable and reliable results. A threshold (e.g., 0.5) is then applied to the probability map output by the model to generate a binary segmentation mask. If post-processing is required, such as eliminating isolated small regions or smoothing edges, morphological filtering can also be performed on the segmentation mask at this stage to optimize the result quality.

[0034] Step S5: Result Output and Visualization. The obtained tumor segmentation results are fused and displayed with the original MRI image. For example, the segmented tumor regions can be highlighted with different colors and superimposed on a grayscale MRI, allowing doctors to visually verify the segmentation effect. Simultaneously, based on clinical needs, the two-dimensional segmentation results of continuous slices can be reconstructed into a three-dimensional tumor volume map. Doctors can observe the three-dimensional tumor morphology from different angles and perform a comprehensive evaluation in conjunction with the original image. Finally, the system outputs segmented images of head and neck tumors and related data (such as quantitative information like tumor volume and maximum diameter) for medical diagnosis and treatment planning.

[0035] To evaluate the multi-task classification performance of GTVp and GTVn, this work uses three evaluation metrics: Dice similarity coeddicient (Dice), intersection over union (IOU), and Hausdorff distance (HD). The relevant calculation methods are as follows: Among them, true positive (TP) represents the number of actual lesion area samples successfully identified by the model, reflecting the model's ability to capture target areas; true negative (TN) represents the number of background area samples correctly classified, reflecting the model's ability to exclude non-lesion areas; false positive (FP) represents the number of samples that were mistakenly identified as lesions; and false negative (FN) represents the number of actual lesion samples that were not identified, reflecting the degree of omission in segmentation.

[0036] exist During the calculation, This is the segmentation result image. This is the true label image. Considering that some outliers can interfere with the segmentation result evaluation, we selected the top [label] points during the actual calculation. Distance values. This embodiment compares images annotated manually, using existing models, and annotated with the algorithm of this patent. The results are as follows: Figure 1 As shown.

[0037] Table 1. DSC Evaluation Table for Different Algorithms After gradually introducing the aforementioned innovative modules in the experiments, the model segmentation accuracy was significantly improved. Compared to the basic U-Net model, the average Dice of the model after introducing the spatial-channel attention module (such as "SCU-Net") increased to approximately 79.00%, the average Dice of the model after introducing the dense hollow pooling module (such as "DenseU-Net") was approximately 80.00%, while the average Dice of the model of this invention, which simultaneously employs attention and dense hollow pooling, reached approximately 85.20%. This result demonstrates that the combination of the modules designed in this invention can synergistically improve the accuracy of tumor segmentation.

[0038] Table 2 Evaluation of the effects of different algorithm techniques This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. An MRI head and neck tumor segmentation system based on the SCDU-Net model, characterized in that: The system includes an encoder module, a bottleneck module, and a decoder module. The encoder module has multi-level downsampling, with each downsampling level containing a residual convolutional block and a coordinate attention module connected to the residual convolutional block, used to extract multi-level features of the image and progressively reduce the spatial resolution. The bottleneck module is connected to the last downsampling layer of the encoder module and contains a densely dilated spatial pyramid pooling module, used to extract multi-scale contextual information of the target region through multiple dilated convolutional branches with different dilation rates. The decoder module is connected to the bottleneck module and receives skip connection features from each downsampling layer of the encoder module. The decoder module has multi-level upsampling, used to progressively restore the spatial resolution of the feature map and output a segmentation result image with the same size as the input image.

2. The MRI head and neck tumor segmentation system based on the SCDU-Net model according to claim 1, characterized in that: The residual convolutional block includes at least two convolutional layers and a residual connection that adds the input of the residual convolutional block to the output of the convolutional layers, which is used to enhance feature extraction capabilities without increasing network depth.

3. The MRI head and neck tumor segmentation system based on the SCDU-Net model according to claim 1, characterized in that: The coordinate attention module extracts directional position information encoding by performing global pooling along the width and height directions of the input feature map, generating corresponding horizontal and vertical attention weight maps, and enhancing spatial attention on the input feature map. The output of the coordinate attention module is fused with the output of the one-channel attention SE module to simultaneously calibrate the feature map spatially and channel-wise.

4. The MRI head and neck tumor segmentation system based on the SCDU-Net model according to claim 1, characterized in that: The densely hollow spatial pyramid pooling module includes at least three hollow convolutional layers with different dilation rates. The input of each hollow convolutional layer includes the original input feature map of the densely hollow spatial pyramid pooling module and the output features of all previous hollow convolutional layers with dilation rates less than that of the hollow convolutional layer. The receptive field is expanded and multi-scale contextual information is obtained through layer-by-layer dense connection.

5. The MRI head and neck tumor segmentation system based on the SCDU-Net model according to claim 1, characterized in that: The encoder module also includes a convolutional coding branch and a Transformer coding branch in parallel, used to extract local features and long-range dependency features of the input image; the features extracted by the Transformer coding branch are fused with the features of the convolutional coding branch in the decoder module through an attention mechanism to enrich the global context information in the decoding stage.

6. The MRI head and neck tumor segmentation system based on the SCDU-Net model according to claim 1, characterized in that: The decoder module has at least two different levels of upsampling layers connected to a segmentation output submodule, which is used to output intermediate segmentation result images at corresponding resolutions to provide multi-scale supervision signals during model training.

7. A segmentation method for an MRI head and neck tumor segmentation system based on the SCDU-Net model according to any one of claims 1-6, characterized in that, Includes the following steps: Step 1, Image Preprocessing: The original MRI image is intensity normalized and resampled to a predetermined uniform resolution to generate a single-channel two-dimensional image; random data augmentation processing is performed on the image, including random morphological transformation, random geometric transformation and random intensity transformation, and image slices that do not contain tumor regions are filtered out. Step 2, Feature Extraction: The preprocessed image is input into the encoder module of the segmentation system. Multi-level features are extracted through residual convolutional blocks and spatial information is enhanced by the coordinate attention module. After multi-level downsampling, a low-resolution deep feature map is obtained. Step 3, Context Feature Acquisition: Input the low-resolution feature map obtained in Step 2 into the densely dilated spatial pyramid pooling module of the bottleneck module, extract and fuse context features of different scales through dilated convolution branches with multiple dilation rates to obtain enhanced feature representations. Step 4, Upsampling Decoding: Input the feature representation obtained in Step 3 into the decoder module of the segmentation system, upsample step by step to restore spatial resolution, and fuse the skip connection features from the corresponding encoder layer to generate a segmentation result image with the same size as the original image; Step 5, Loss Calculation and Model Update: The segmentation result image is compared with the corresponding ground truth label at the pixel level. The combined loss value consisting of Dice loss and Focal loss is calculated, and the model parameters of the segmentation system are updated and adjusted based on the loss value through the backpropagation algorithm.

8. The segmentation method according to claim 7, characterized in that: In step 4, the decoder module outputs intermediate segmentation result images of corresponding resolutions in multiple upsampling stages, calculates the loss value between each intermediate segmentation result image and the corresponding scaled real label, and finally weights and sums the loss values ​​of each resolution to obtain the total loss, thereby realizing multi-scale deep supervised training.

9. The segmentation method according to claim 7, characterized in that: The calculation of the loss value further includes introducing a boundary loss term, which gives higher weight to the error at the edge of the target region in the segmentation result image.

10. The segmentation method according to claim 7, characterized in that: In step 2, the image features are extracted in parallel through a convolutional encoder branch and a Transformer encoder branch, wherein the Transformer encoder branch is used to extract long-range dependency features of the image. In step 4, the features extracted by the Transformer encoder branch and the features of the convolutional encoder branch are fused together through an attention mechanism and then decoded to obtain a more complete segmentation result image.

Images