A three-dimensional medical image segmentation method based on CircularUMamba

Abstract

The application discloses a three-dimensional medical image segmentation method based on CircularUMamba, and belongs to the technical field of medical image analysis and deep learning. The method comprises the following steps: acquiring 3D medical image data as a training set, and performing pretreatment and preliminary feature extraction on the image data; constructing a backbone segmentation network based on CircularUMamba, processing the features output by S1 by using the network, and finally outputting a multi-level voxel-level segmentation prediction probability map; training the network constructed by S2 by using a combined loss function, optimizing the network parameters by minimizing the total training loss, and obtaining a trained image segmentation model; and using the trained model to obtain a voxel-level segmentation prediction probability map as a segmentation result for an image to be segmented. The application can improve the recognition accuracy of micro lesions and fuzzy tissue boundaries.

Description

Technical Field

[0001] This invention relates to the fields of medical image analysis and deep learning technology, specifically to a three-dimensional medical image segmentation method based on CircularUMamba, applicable to the segmentation of organs and lesions in CT or MRI images. Background Technology

[0002] Medical image segmentation is a crucial step in clinical workflows such as computer-aided diagnosis (CAD), surgical planning, and radiotherapy. Its core lies in accurately locating and delineating organ or lesion regions from images such as CT or MRI. Although advancements in deep learning technology have led to increasingly sophisticated automated segmentation models that have gradually replaced tedious manual delineation, significantly improving diagnostic efficiency, the inherent complexity of anatomical structures, the variability in lesion scale, and the high degree of continuity in three-dimensional space within medical images still pose significant challenges to existing algorithms.

[0003] To address the problem of medical image segmentation, numerous solutions have been proposed by academic and industrial communities both domestically and internationally. Xing (Xing, Z., Ye, T., Yang, Y., Liu, G., Zhu, L. (2024). SegMamba: Long-Range Sequential Modeling Mamba for 3D Medical Image Segmentation. In: Linguraru, MG, et al. The Medical Image Computing and Computer Assisted Intervention – MICCAI 2024 proposed SegMamba, which aims to leverage the Mamba architecture's advantage in handling long sequences by flattening 3D voxel data into one-dimensional sequences and combining this with a multi-directional scanning strategy to simulate 3D global context. However, this operation, which forcibly maps voxels with their original three-dimensional spatial structure into independent one-dimensional sequences, fundamentally disrupts the inherent spatial topological continuity between voxels. Adjacent voxels may be far apart after serialization, leading to severe loss of contextual information when the model captures local geometric features. Wang, J., Chen, J., Chen, D., Wu, J. (2024). LKM-UNet: Large Kernel Vision Mamba UNet for Medical Image Segmentation. In: Linguraru, MG, et al.The Medical Image Computing and Computer-Assisted Intervention (MICCAI 2024) project attempted to optimize feature extraction by introducing a hybrid design of large-kernel convolution and Mamba. This method allocates a larger spatial window to the SSM module to expand the receptive field and uses pooling operations to aggregate and compress features within the window before feeding them into Mamba for scanning. While this design attempts to balance global information with local features, the window-based partitioning mechanism easily introduces artificial boundary effects in non-overlapping regions. More importantly, the reliance on pooling layers to integrate features inevitably smooths out originally sharp edge information, sacrificing crucial fine-grained structural details in medical image segmentation, thus limiting the model's final segmentation accuracy.

[0004] In summary, current medical image segmentation models have the following shortcomings: (1) Existing global scanning methods break the inherent orthogonal neighborhood relationship and spatial locality between voxels when converting 3D data into 1D sequences, making it difficult for the model to effectively maintain the original 3D geometric structure features in long sequences.

[0005] (2) Although existing window-based methods focus on locality, they lack effective cross-window interaction mechanisms, which leads to interruption of information flow at window boundaries and produces boundary artifacts. Furthermore, some methods lose high-frequency detail information during feature aggregation, which limits the final performance of the model in segmenting complex anatomical structures.

[0006] Accurate and automatic segmentation of 3D medical images is a cornerstone technology for quantitative disease analysis, surgical planning, and radiotherapy. The reliability of the segmentation results directly affects the safety and effectiveness of clinical diagnosis and treatment. However, current mainstream segmentation methods suffer from two major technical bottlenecks: First, traditional deep learning models struggle to balance global context and computational efficiency. CNN architectures are limited by local receptive fields, making it difficult to capture long-distance dependencies; while Transformer architectures can model global information, their quadratic computational complexity leads to high resource consumption, making it difficult to efficiently process high-resolution 3D images. Second, existing Mamba improvements disrupt the integrity of 3D spatial structures. Whether it's simple global sequential scanning or rigid windowing strategies, both inevitably sever the spatial neighborhood relationships between voxels or introduce artificial boundary artifacts, limiting the model's segmentation accuracy for small lesions and the edges of complex anatomical structures. Summary of the Invention

[0007] To overcome the shortcomings of existing technologies, this invention proposes a three-dimensional medical image segmentation method based on CircularUMamba, comprising the following steps: Step 1: Obtain 3D medical image data as a training set, and input the image data into the Stem layer for preliminary feature extraction; Step 2: Construct a backbone segmentation network based on CircularUMamba, which includes an encoder module, a residual module, and a decoder module. Use this network to process the features output by the Stem layer. The encoder module uses the Circular Mamba module and the TSM operator to perform multi-level feature extraction. The network finally outputs a multi-level voxel-level segmentation prediction probability map. Step 3: Based on the features obtained in Step 1, the network constructed in Step 2 is trained using a combined loss function. The network parameters are optimized by minimizing the total training loss to obtain a trained image segmentation model. Step 4: For the 3D medical image data to be segmented, the preliminary features extracted in Step 1 are used. The preliminary features are then input into the image segmentation model trained in Step 3 to obtain a voxel-level segmentation prediction probability map as the segmentation result.

[0008] According to a preferred embodiment of the present invention, step 1, in addition to performing preliminary feature advance, also includes a step of preprocessing the image data. The preprocessing in S1 includes one or more of the following: data dimension adjustment, coordinate system adjustment, voxel spacing adjustment, intensity normalization, background content cropping, and foreground content filling.

[0009] According to a preferred embodiment of the present invention, in S1, a Stem layer is used for preliminary feature extraction. The Stem layer consists of a three-dimensional convolutional operation with a kernel size of 7 × 7 × 7, a stride of 2, and a padding of 3. It performs spatial downsampling and feature dimension enhancement operations on the image data, halving the image size and expanding the grayscale channels into multi-channel feature maps.

[0010] Furthermore, the specific structure of the backbone segmentation network described in step 2 is as follows: the backbone segmentation network is a multi-level structure, defined sequentially from the first layer to the Nth layer in the depth direction; i is defined as the level index; The first layer takes the original image as input directly. This layer does not have an encoder module. After the original image is processed by the residual module, it is directly used as a skip connection feature and input to the decoder module. The output of the decoder module in this layer is connected to a segmentation head. The second layer has a Stem module as its encoding end. The output features are directly input to the residual module of the current layer and also serve as the input to the next layer. The output of the decoder module of this layer is connected to the deep supervision module. From layer 3 to layer N-1, the encoders are all composed of cyclic Mamba modules. The output features are directly input to the residual module of the current layer and used as the input of the next layer. The output of each layer's decoder module is connected to a deep supervision module. The Nth layer's encoder consists of a cyclic Mamba module. This layer does not have a decoder module. Its output is directly transmitted to the decoder module of the (N-1)th layer after passing through the residual module and upsampling operation. The encoder module in the backbone segmentation network uses a recurrent Mamba module to perform deep feature extraction on the input features, and then performs resolution compression through a downsampling layer to obtain the output. The output of the previous level encoder module is used as the input of the next level encoder module, thereby realizing multi-level feature extraction. The residual module performs convolution operations to adjust the channel dimensions and extract high-frequency spatial detail information, outputting skip connection features. ; The i-th layer decoder module performs an upsampling operation to restore the resolution of the (i+1)-th layer features to match that of the current layer. It then concatenates the upsampled features with the skip connection features along the channel dimension. Subsequently, a cyclic Mamba module is used for feature fusion and decoding reconstruction to output the features. ; The deep supervision module is used to map the features of the intermediate layers to auxiliary segmentation results and calculate auxiliary loss to assist network training. The segmentation head is used to map the number of feature channels to the number of target categories through convolutional layers to generate the final voxel-level segmentation results.

[0011] The specific execution process of the backbone segmentation network is as follows: Step 21: Its input features The definition is as follows: ; In the formula, These are the initial features output by the Stem layer. This is the output of the encoder module in the previous layer; The encoder module first uses a recurrent Mamba module to perform deep feature extraction on the input features, and then performs resolution compression through a downsampling layer to obtain the final output of the current layer; the input-output relationship of this process is defined by the following formula: ; In the formula, This indicates the feature extraction operation of the looping Mamba module; DownSampling indicates downsampling.

[0012] The looping Mamba module consists of three sub-modules connected in series: window Mamba, looping window Mamba, and global Mamba. The window Mamba sub-module is denoted as WM, the looping window Mamba sub-module as CWM, and the global Mamba sub-module as GM. The specific execution process includes the following steps: S211: WM stage: Input features Perform a window partitioning operation to reshape it into a window containing A set of non-overlapping local 3D window features ; ; In the formula, Represents window partitioning operations. Representing the The feature tensor of a window For window size, For the number of channels, Total number of windows; For sets Each window feature in the dataset is processed using the TSM operator to obtain the processed window feature set. Finally, for the set Perform the inverse window partitioning operation to combine all processed window features back to the dimensions of the original feature map, obtaining the output features of the WM stage. ; S212: CWM Phase: Output of the WM phase Perform a cyclic shift operation in the spatial dimension, with a shift step size. The displacement feature is obtained by setting the window size to half its original size. ;right Perform the same window partitioning operation as in S211, and apply the TSM operator defined above to each window; after the operation is completed, perform the inverse window partitioning and inverse circular shift operations in sequence to restore the spatial position of the feature map and obtain the output feature. ; S213: GM stage: Retrieves the complete features output by the CWM stage. The data is directly used as input without window partitioning, and the TSM operator is applied to perform a full-map sequence scan.

[0013] Preferably, the specific construction and operation process of the TSM operator is defined as follows: The input features of the TSM operator are First, a convolution kernel with a size of The deep convolution process obtains the input features : ; In the formula, Represents depthwise convolution; Will The inputs are fed into three independent Mamba modules, and sequence scans are performed along three orthogonal directions: forward, backward, and interlayer, to obtain the scan output features. ; ; ; ; In the formula, , and These represent Mamba modules that employ forward, reverse, and inter-layer scanning strategies, respectively. Finally, It is concatenated with the three scan output features along the channel dimension, and then... Fusion convolutional layers Perform feature aggregation to obtain the TSM output. : .

[0014] According to a preferred embodiment of the present invention, the specific execution process of the i-th layer residual module includes: Step 221: Perform convolution, normalization, and activation operations sequentially on the input features of the input residual module, followed by convolution and normalization operations to obtain the residual mapping features. : Step 222: Map the residual features output from step 221. The original input features are added element-wise, and the Leaky ReLU activation function is applied to the result to obtain the final output of the i-th layer residual module. .

[0015] According to a preferred embodiment of the present invention, the execution process of the i-th layer decoder module includes: Step 231: Upsample the output features from the (i+1)th decoder module by deconvolution, doubling the spatial size of the feature map to match the resolution of the skip connection features at the same level, thus obtaining the upsampled features. : Step 232: Upsample the features Skip connection characteristics with the output of the residual module The features are stitched together along the channel dimension to obtain the fused features. : Step 233: Merge features The input is fed into the decoding convolutional block to obtain the output features of the decoder. .

[0016] According to a preferred embodiment of the present invention, the combined loss function in S3 Defined as: ; In the formula, For the first Weighting coefficients for hierarchical loss; For the first Single-layer loss function

[0017] ; The cross-entropy loss function, The Dice loss function; As the loss balance factor; Cross-entropy loss The calculation is as follows: For each output level First, the prediction map of this level Upsampled to the actual label The same resolution; subsequently, cross-entropy loss is calculated based on voxel-level classification accuracy. : ; In the formula, The total number of voxels; The number of categories; As an indicator function, when voxels Category The value is 1 if it is true, and 0 otherwise. To predict probabilities; Dice loss function The calculation is as follows: ; In the formula, This indicates the number of elements in the intersection of the predicted graph and the true labels; and These represent the number of elements in the predicted image and the actual label, respectively.

[0018] Compared with the prior art, the beneficial effects of the present invention include: This invention constructs an intelligent solution for 3D medical image segmentation tasks by introducing a hierarchical feature modeling paradigm from local to global. It innovatively introduces a spatial cyclic shift mechanism into the Mamba architecture, enabling the model to effectively eliminate boundary artifacts caused by traditional block processing through fine-grained feature extraction within windows and seamless interaction across windows, while maintaining linear computational complexity. This also ensures both local structural integrity and global context awareness. Furthermore, by combining tri-orthogonal scanning (TSM) with a multi-scale window cascading mechanism, it can perform comprehensive analysis of 3D volumetric data, constructing isotropic feature representations. This effectively corrects feature biases caused by a single scanning direction, further improving the recognition accuracy of minute lesions and blurred tissue boundaries. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the backbone segmentation network architecture of the 3D medical image segmentation method based on CircularUMamba provided in this embodiment of the invention.

[0020] Figure 2 This is a detailed structural block diagram of the three-dimensional spatial Mamba (TSM) module in this invention.

[0021] Figure 3 The results show a qualitative comparison of the segmentation performance of each method on the BTCV dataset.

[0022] Figure 4 Grad-CAM diagrams for SegMamba and CircularUMamba models. Detailed Implementation

[0023] The following detailed implementation of the Mamba-based 3D medical image segmentation method of this invention will be illustrated with specific examples. The steps are as follows: Step 1: Acquire 3D medical image data. In this embodiment, the medical image data uses the BTCV multi-organ segmentation dataset, with 24 images used as the training set and 6 images used as the test set. The image data undergoes a preprocessing procedure, including data dimension adjustment, coordinate system adjustment, voxel spacing adjustment, intensity normalization, background content cropping, and foreground content filling. Then, it is input into the Stem layer for preliminary feature extraction. In this embodiment, 3D medical image data is acquired. Where B, C, D, H, and W represent batch size, number of channels, depth, height, and width, respectively. In this embodiment, B, C, D, H, and W are 4, 1, 64, 64, and 64, respectively. The Stem layer consists of a three-dimensional convolutional kernel with a size of [missing information]. The output consists of convolution operations with a stride of 2 and padding of 3, producing an initial high-dimensional feature map. ,in This represents the initial feature embedding dimension after projection through the Stem layer. In this embodiment, It is 48; Step 2: Construct a backbone segmentation network based on CircularUMamba. This network has a multi-level structure, defined sequentially as layers 1 to N in the depth direction. The first layer directly takes the original image as input. This layer does not have an encoder module. After processing by the residual module, the original image is directly used as a skip connection feature and input to the decoder module. The output of the decoder module of this layer is connected to the segmentation head. The second layer has a Stem module as its encoder. The output feature is directly input to the residual module of the current layer and simultaneously used as the input to the next layer. The output of the decoder module of this layer is connected to the depth supervision module. The encoders of layers 3 to N-1 are all composed of cyclic Mamba modules. The output feature is directly input to the residual module of the current layer and used as the input to the next layer. The output of the decoder module of each layer is connected to the depth supervision module. The encoder of layer N is composed of cyclic Mamba modules. This layer does not have a decoder module. Its output is directly transmitted to the decoder module of layer N-1 after processing by the residual module and upsampling operation.

[0024] The network input is The output is a segmentation prediction probability map. The encoder module in the backbone segmentation network uses a recurrent Mamba module to perform deep feature extraction on the input features, and then performs resolution compression through a downsampling layer to obtain the output. The output of the previous level encoder module is used as the input of the next level encoder module, thereby realizing multi-level feature extraction. The residual module performs convolution operations to adjust the channel dimensions and extract high-frequency spatial detail information, outputting skip connection features. ; The decoder module of layer i performs an upsampling operation to restore the resolution of the features of layer i+1 to be consistent with the current layer. It then concatenates the upsampled features with the skip connection features along the channel dimension. Subsequently, a cyclic Mamba module is used for feature fusion and decoding reconstruction to output the features. ; The deep supervision module is used to map the features of the intermediate layers to auxiliary segmentation results and calculate auxiliary loss to assist network training. The segmentation head is used to map the number of feature channels to the number of target categories through convolutional layers to generate the final voxel-level segmentation results.

[0025] like Figure 1 As shown, a 6-level structure is given in this embodiment; defined This is a hierarchical index; the specific construction process is as follows: Step 21: For the first The encoder of each layer (encoders are configured in layers 3-6 of the backbone segmentation network), its input features The definition is as follows:

[0026] In the formula, These are the initial features output by the Stem layer. This is the output of the previous layer encoder module. This module first uses a recurrent Mamba module to perform deep feature extraction on the input features, and then performs resolution compression through a downsampling layer to obtain the final output of the current layer. The input-output relationship of this process is defined by the following unified formula: ; In the formula, This indicates the feature extraction operation of the recurrent Mamba module; DownSampling indicates the use of a convolution kernel size of... Downsampling is achieved through a convolution operation with a stride of 2; Furthermore, the aforementioned The operation consists of three sub-modules: window Mamba, loop window Mamba, and global Mamba, executed sequentially. The specific execution process includes the following steps: Step 211: First, analyze the input features. Perform a window partitioning operation to reshape it into a window containing A set of non-overlapping local 3D window features Its definition is as follows:

[0027] In the formula, Representing the The feature tensor of a window For the window size, this embodiment employs a multi-scale cascading strategy, setting it to 4 and 8 sequentially for serial processing. For the number of channels, Total number of windows; Next, for the set Each window feature in the dataset is processed using the three-dimensional space Mamba (TSM) operator to obtain the processed window feature set. :

[0028] Finally, for the set Perform the inverse window partitioning operation to combine all processed window features back to the dimensions of the original feature map, obtaining the output features of the WM stage. :

[0029] In the formula, This represents a reverse window partitioning operation; The specific construction and operation process of the TSM operator is defined as follows: The input features of TSM are First, a convolution kernel with a size of The deep convolution process obtains the input features :

[0030] In the formula, Represents depthwise convolution; Will The inputs are fed into three independent Mamba modules, and sequence scans are performed along three orthogonal directions: forward, backward, and interlayer, to obtain the scan output features. ; ; ; ; In the formula, , and These represent Mamba modules that employ forward, reverse, and inter-layer scanning strategies, respectively. Finally, It is concatenated with the three scan output features along the channel dimension, and then... Fusion convolutional layers Perform feature aggregation to obtain the TSM output. : .

[0031] Step 212: Output of the WM stage Perform a cyclic shift operation in the spatial dimension, with a shift step size. The displacement feature is obtained by setting the window size to half its original size. ;right Perform the same window partitioning operation as in S211, and apply the TSM operator defined above to each window; after the operation is completed, perform the inverse window partitioning and inverse circular shift operations in sequence to restore the spatial position of the feature map and obtain the output feature. ; S213: GM stage: Retrieves the complete features output by the CWM stage. The data is directly used as input without window partitioning, and the TSM operator is applied to perform a full-map sequence scan.

[0032] Step 22: The specific execution process of the i-th layer residual module includes: Step 221: Perform convolution, normalization, and activation operations sequentially on the input features of the input residual module, followed by convolution and normalization operations to obtain the residual mapping features. : Step 222: Map the residual features output from step 221. The original input features are added element-wise, and the Leaky ReLU activation function is applied to the result to obtain the final output of the i-th layer residual module. .

[0033] In one specific embodiment of the present invention, for any residual module of layer i in layers 3 to 6, its input is the output of the encoder of layer i. The input-output relationship of this process is defined by the following unified formula:

[0034] In the formula, This indicates the processing operations of the residual module. This is the output characteristic of this module. The specific execution process includes the following sub-steps: Step 221: Process the input features The convolution operation, normalization, and activation operation are performed sequentially, followed by another convolution operation and normalization operation to obtain the residual mapping features. :

[0035]

[0036] In the formula, and Each represents two independent convolutional layers, both with a kernel size of [missing information]. Convolution operations; and For the corresponding normalization operation; Use the Leaky ReLU activation function; It represents an intermediate transitional feature.

[0037] Step 222: Map the residual features output from step 221. and original input features Element-wise addition is performed, and the Leaky ReLU activation function is applied to the sum to obtain the final output of the i-th layer residual module. :

[0038] In the formula, This represents the Leaky ReLU activation function.

[0039] Step 23: The input to the i-th decoder module consists of two parts: features from the decoder at layer (i+1). And the skip connection features from the current level residual module. It should be noted that since the Nth layer (N=6 in this embodiment) does not have a decoder module, the decoder module of the (N-1)th layer receives the features of the Nth layer residual module and the upsampling operation, as well as the skip connection features from the (N-1)th layer residual module; the specific execution process includes the following sub-steps: Step 231: Upsample the features from the (i+1)th decoder by deconvolution, doubling the spatial size of the feature map to match the resolution of the skip connection features at the same level, thus obtaining the upsampled features. :

[0040] In the formula, This represents a resolution restoration operation with an upsampling factor of 2.

[0041] Step 232: Perform feature concatenation. Combine the upsampled features... Features of skip connections The features are stitched together along the channel dimension to obtain the fused features. :

[0042] In the formula, This represents the splicing operation of the channel dimension.

[0043] Step 233: Analyze the fusion features Decoding is then performed. The input is fed into the decoding convolutional block to obtain the output features of the current i-th layer decoder. :

[0044] In the formula, This indicates the residual processing operation applied during the decoding stage, whose internal operation logic is the same as defined in step 22. To maintain consistency, only the number of input and output channels is adapted according to decoding requirements.

[0045] Step 24: [To be continued] The kernel size is [size to be filled in] The convolution operation maps the features to the target classification space, resulting in the ... Hierarchical prediction graph :

[0046] In the formula, Indicates by The classification mapping unit, constructed by convolution, has an output channel number equal to the number of target categories. For the first level, the classification mapping unit is the segmentation head, and the resulting... This is the final segmentation prediction map; for levels 2 to N-1, the classification mapping unit is the deep supervision module, resulting in... This is an auxiliary prediction graph.

[0047] Step 3: Train the model using a combined loss function. Optimize the network parameters by minimizing the total training loss to obtain a trained image segmentation model.

[0048] Furthermore, step 3 specifically includes the following steps: Step 31: Calculate the cross-entropy loss function ; For each output level First, the prediction map of this level Upsampled to the actual label Same resolution. Then, cross-entropy loss is calculated based on voxel-level classification accuracy. :

[0049] In the formula, The total number of voxels; The number of categories; As an indicator function, when voxels Category The value is 1 if it is true, and 0 otherwise. To predict probabilities; Step 32: Calculate the Dice loss function ; To address the class imbalance problem and assess the overlap between predicted and true regions, the Dice loss is calculated. :

[0050] In the formula, This indicates the number of elements in the intersection of the predicted graph and the true labels; and These represent the number of elements in the predicted image and the true label, respectively. Step 33: Calculate the total loss of a single layer; Cross-entropy loss function and Dice loss function Weighted combinations constitute the single-layer loss function for model training. ;

[0051] In the formula, The loss balance factor is set to 1.0 in this embodiment.

[0052] Step 34: Calculate the total training loss of the network; For all The total training loss is obtained by weighted summing of the total losses of each output layer. And optimize all learnable parameters using the gradient descent algorithm;

[0053] In the formula, For the first In this embodiment, the weighting coefficients for the hierarchical loss are set to 8 / 15, 4 / 15, 2 / 15, and 1 / 15 respectively.

[0054] To verify the effectiveness of the medical image segmentation model in this embodiment of the invention, experiments were conducted using the BTCV dataset and the MSD-Liver dataset, and the model was compared with several existing medical image segmentation models.

[0055] The BTCV dataset is a widely used benchmark dataset for segmentation research of multiple abdominal organ CT images. It contains 30 CT scan images covering 13 human organs: spleen, right kidney, left kidney, gallbladder, esophagus, liver, stomach, aorta, inferior vena cava, portal vein and splenic vein, pancreas, right adrenal gland, and left adrenal gland. In this study, the standard segmentation method for the BTCV dataset by Swin UNETR was followed: 24 images were used for model training, while the remaining 6 images were used for performance testing. The Dice coefficients for each of the 13 organs are reported to evaluate the accuracy of the segmentation model.

[0056] The Liver Tumor Dataset in MSD is part of the internationally recognized Medical Segmentation Decathlon and a widely used benchmark dataset in the field of liver and liver tumor CT image segmentation. This dataset contains 131 3D contrast-enhanced abdominal CT scans, collected from patients with primary or metastatic liver disease caused by colorectal cancer, breast cancer, or lung cancer. Each scan provides fine pixel-level annotations for two target categories: liver and liver tumor. A significant challenge of this dataset is the severe label imbalance problem—which places higher demands on the robustness and accuracy of segmentation models. Therefore, this dataset was chosen to evaluate model performance. For model training and evaluation, the image data was randomly divided into training, validation, and test sets in a 70:15:15 ratio.

[0057] This embodiment uses two key metrics, Dice and NSD, to comprehensively evaluate image segmentation quality. The Dice coefficient quantifies the volumetric similarity between the segmentation result and the corresponding ground truth label, reflecting the spatial overlap between the predicted and ground truth regions. The NSD metric quantifies the boundary-level consistency between the predicted result and the ground truth label. This metric introduces a spatial tolerance threshold to calculate the degree of surface overlap between the predicted and ground truth contours within an allowable error range, thereby evaluating the model's segmentation accuracy for anatomical structure edge details. The calculation formulas for the two metrics are shown below:

[0058] In the formula, TP, FP, and FN represent the number of pixels (or voxels) that are true positives, false positives, and false negatives, respectively.

[0059]

[0060] In the formula, To predict segmentation regions, This is the actual label area; and They are respectively and Surface point set; and They are respectively and The point in the middle; Indicates the number of points in the set; Representative point to the set The shortest Euclidean distance. In this embodiment, this threshold is... Set to 5 mm.

[0061] Table 1. Quantitative segmentation results on the BTCV dataset (best results are shown in bold).

[0062] Table 1 shows the Dice coefficient results of the CircularUMamba model and existing medical image segmentation models on the BTCV dataset in this embodiment of the invention. The model's average Dice coefficient reaches 84.55%, comprehensively surpassing existing state-of-the-art methods. This performance is not only 0.57% higher than nnUNet (83.98%), but also 3.11% higher than SwinUNETR (81.44%) based on the Transformer architecture, and 1.55% higher than the similar state-of-the-art method SegMamba.

[0063] Table 2. Quantitative segmentation results on the MSD-Liver dataset (best results are shown in bold).

[0064] Table 2 shows the Dice and NSD coefficient results of the CircularUMamba model and existing medical image segmentation models in this embodiment of the invention on the MSD-Liver dataset. Our model ranks first in both the average Dice (78.64%) and average NSD (84.70%), comprehensively surpassing cutting-edge methods including nnUNet and SegMamba. Our model performs exceptionally well on the boundary-sensitive NSD metric, achieving a score as high as 74.95% in tumor segmentation.

[0065] Table 3 shows the ablation experiments performed on the proposed module on the BTCV dataset.

[0066] To systematically verify the effectiveness of the core designs in the model, a series of detailed ablation experiments were conducted on the BTCV dataset, and the results are shown in Table 3. The results show that introducing windowed Mamba (WM) or cyclic windowed Mamba (CWM) alone can bring performance improvements of 0.52% and 0.35%, respectively. Notably, the model performance reached 83.15% when using only global Mamba (GM), achieving the largest single-item gain (+0.78%). Next, the synergistic effect of the two local modeling strategies was explored. When WM and CWM were combined, the model performance significantly improved to 83.44%. Finally, the integration effect of all components was verified. After integrating all three core components—WM, CWM, and GM—the best Dice score of 84.55% was achieved.

[0067] Table 4 "4+8" represents the proposed cascaded multi-scale strategy.

[0068] Window configuration is a key design element in the model. To explore the optimal windowing strategy, a series of experiments were conducted to determine the best window size, and the results are shown in Table 4 (the strategy in Table 4 consists of a 4×4×4 window module followed by an 8×8×8 window module). The experiments first evaluated the impact of a single fixed window size (2, 4, 8). The results show that, at a single scale, the model performs best (84.08%) with a window size of 4, outperforming configurations with sizes of 2 (83.87%) and 8 (83.70%). Building on this, we further introduced a "4+8" cascaded multi-scale strategy. Experimental data show that this strategy effectively surpasses all single-window configurations, achieving the highest Dice score of 84.55%.

[0069] Table 5 Ablation studies on the scanning direction of the TSM module

[0070] Table 5 shows the ablation experimental results for the three scanning directions and their different combinations in the TSM module.

[0071] The experiment first evaluated the performance of a single scan direction, with reverse scanning achieving the highest Dice score of 83.94%. Subsequently, the experiment tested the effect of bidirectional scan combinations, showing that bidirectional combinations generally outperformed single-direction scans, with the combination of forward and reverse (Forward + Reverse) achieving a Dice score of 84.21%. The final results indicate that the model achieves optimal performance, reaching a maximum Dice score of 84.55%, when integrating forward, reverse, and inter-layer scan strategies simultaneously.

[0072] Figure 3 A qualitative comparison of the segmentation performance of various methods on the BTCV dataset is provided. It can be observed that for organs with high segmentation difficulty, such as the stomach and pancreas, many comparative methods struggle to achieve accurate contour delineation. In contrast, CircularUMamba demonstrates comprehensive performance advantages in these challenging scenarios. Specifically, as shown in rows 1, 2, 3, and 5, even when dealing with large stomach structures, the model still exhibits excellent robustness, and its segmentation results closely match the ground truth (GT).

[0073] Figure 4Grad-CAM images of the SegMamba and CircularUMamba models are provided. The visualization results clearly show that CircularUMamba produces more accurate and stronger activation in local regions. For example, in the third column, regarding the activation map of the inferior vena cava (IVC), the SegMamba activation map is not only weaker, but its spatial coverage also fails to accurately match the target. In contrast, the activation map generated by CircularUMamba is significantly stronger and accurately delineates the complete outline of the IVC.

[0074] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A three-dimensional medical image segmentation method based on CircularUMamba, characterized in that, Includes the following steps: S1: Obtain 3D medical image data as a training set and perform preliminary feature extraction on the image data; S2: Construct a backbone segmentation network based on CircularUMamba, which includes an encoder module, a residual module, and a decoder module; use this network to process the features output from S1, and finally output a multi-level voxel-level segmentation prediction probability map; among them, the encoder module uses the Circular Mamba module and the TSM operator to perform multi-level feature extraction. S3: The network constructed in S2 is trained using a combined loss function. The network parameters are optimized by minimizing the total training loss to obtain a well-trained image segmentation model. S4: For the 3D medical image data to be segmented, preliminary features are extracted using S1 and input into the image segmentation model trained by S3 to obtain a voxel-level segmentation prediction probability map as the segmentation result.

2. The three-dimensional medical image segmentation method based on CircularUMamba according to claim 1, characterized in that, In addition to the initial feature extraction, the process also includes a preprocessing step for the image data. This preprocessing includes one or more of the following: data dimension adjustment, coordinate system adjustment, voxel spacing adjustment, intensity normalization, background content cropping, and foreground content filling.

3. The three-dimensional medical image segmentation method based on CircularUMamba according to claim 1, characterized in that, In S1, the Stem layer is used for preliminary feature extraction. The Stem layer consists of a three-dimensional convolutional operation with a kernel size of 7 × 7 × 7, a stride of 2, and padding of 3. It performs spatial downsampling and feature dimension enhancement operations on the image data, halving the image size and expanding the grayscale channels into multi-channel feature maps.

4. The three-dimensional medical image segmentation method based on CircularUMamba according to claim 1, characterized in that, The backbone segmentation network described in S2 is a multi-level structure, which is defined sequentially from the first layer to the Nth layer in the depth direction; The first layer takes the original image as input directly. This layer does not have an encoder module. After the original image is processed by the residual module, it is directly used as a skip connection feature and input to the decoder module. The output of the decoder module in this layer is connected to a segmentation head. The second layer has a Stem module as its encoding end. The output features are directly input to the residual module of the current layer and also serve as the input to the next layer. The output of the decoder module of this layer is connected to the deep supervision module. From layer 3 to layer N-1, the encoders are all composed of cyclic Mamba modules. The output features are directly input to the residual module of the current layer and used as the input of the next layer. The output of each layer's decoder module is connected to a deep supervision module. The Nth layer's encoder consists of a cyclic Mamba module. This layer does not have a decoder module. Its output is directly transmitted to the decoder module of the (N-1)th layer after passing through the residual module and upsampling operation. The encoder module in the backbone segmentation network uses a recurrent Mamba module to perform deep feature extraction on the input features, and then performs resolution compression through a downsampling layer to obtain the output. The output of the previous level encoder module is used as the input of the next level encoder module, thereby realizing multi-level feature extraction. The residual module performs convolution operations to adjust the channel dimensions and extract high-frequency spatial detail information, outputting skip connection features. ; The i-th layer decoder module performs an upsampling operation to restore the resolution of the (i+1)-th layer features to match that of the current layer. It then concatenates the upsampled features with the skip connection features along the channel dimension. Subsequently, a cyclic Mamba module is used for feature fusion and decoding reconstruction to output the features. ; The deep supervision module is used to map the features of the intermediate layers to auxiliary segmentation results and calculate auxiliary loss to assist network training. The segmentation head is used to map the number of feature channels to the number of target categories through convolutional layers to generate the final voxel-level segmentation results.

5. The three-dimensional medical image segmentation method based on CircularUMamba according to claim 4, characterized in that, definition For hierarchical indexes, for the first... The encoder module of the layer, its input features The definition is as follows: ； In the formula, These are the initial features output by the Stem layer. This is the output of the encoder module in the previous layer; The encoder module first uses a recurrent Mamba module to perform deep feature extraction on the input features, and then performs resolution compression through a downsampling layer to obtain the final output of the current layer; the input-output relationship of this process is defined by the following formula: ； In the formula, This indicates the feature extraction operation of the looping Mamba module; DownSampling indicates downsampling.

6. The three-dimensional medical image segmentation method based on CircularUMamba according to claim 5, characterized in that, The looping Mamba module consists of three sub-modules connected in series: window Mamba, looping window Mamba, and global Mamba. The window Mamba sub-module is denoted as WM, the looping window Mamba sub-module as CWM, and the global Mamba sub-module as GM. The specific execution process includes the following steps: S211: WM stage: Input features Perform a window partitioning operation to reshape it into a window containing A set of non-overlapping local 3D window features ; ； In the formula, Represents window partitioning operations. Representing the The feature tensor of a window For window size, For the number of channels, Total number of windows; For sets Each window feature in the dataset is processed using the TSM operator to obtain the processed window feature set. Finally, for the set Perform the inverse window partitioning operation to combine all processed window features back to the dimensions of the original feature map, obtaining the output features of the WM stage. ; S212: CWM Phase: Output of the WM phase Perform a cyclic shift operation in the spatial dimension, with a shift step size. The displacement feature is obtained by setting the window size to half its original size. ;right Perform the same window partitioning operation as in S211, and apply the TSM operator defined above to each window; after the operation is completed, perform the inverse window partitioning and inverse circular shift operations in sequence to restore the spatial position of the feature map and obtain the output feature. ; S213: GM stage: Retrieves the complete features output by the CWM stage. The data is directly used as input without window partitioning, and the TSM operator is applied to perform a full-map sequence scan.

7. The three-dimensional medical image segmentation method based on CircularUMamba according to claim 6, characterized in that, The specific construction and operation process of the TSM operator is defined as follows: The input features of the TSM operator are First, a convolution kernel of size is used. The deep convolution process obtains the input features : ； In the formula, Represents depthwise convolution; Will The inputs are fed into three independent Mamba modules, and sequence scans are performed along three orthogonal directions: forward, backward, and interlayer, to obtain the scan output features. ; ； ； ； In the formula, , and These represent Mamba modules that employ forward, reverse, and inter-layer scanning strategies, respectively. Finally, It is concatenated with the three scan output features along the channel dimension, and then... Fusion convolutional layers Perform feature aggregation to obtain the TSM output. : 。 8. The three-dimensional medical image segmentation method based on CircularUMamba according to claim 4, characterized in that, The specific execution process of the residual module at level i includes: Step 221: Perform convolution, normalization, and activation operations sequentially on the input features of the input residual module, followed by convolution and normalization operations to obtain the residual mapping features. : Step 222: Map the residual features output from step 221 The original input features are added element-wise, and the Leaky ReLU activation function is applied to the result to obtain the final output of the i-th layer residual module. .

9. The three-dimensional medical image segmentation method based on CircularUMamba according to claim 4, characterized in that, The execution process of the i-th layer decoder module includes: Step 231: Upsample the output features from the (i+1)th decoder module by deconvolution, doubling the spatial size of the feature map to match the resolution of the skip connection features at the same level, thus obtaining the upsampled features. : Step 232: Upsample the features Skip connection characteristics with the output of the residual module The features are stitched together along the channel dimension to obtain the fused features. : Step 233: Fuse features The input is fed into the decoding convolutional block to obtain the output features of the decoder. .

10. The three-dimensional medical image segmentation method based on CircularUMamba according to claim 1, characterized in that, S3 describes the combined loss function Defined as: ； In the formula, For the first Weighting coefficients for hierarchical loss; For the first Single-layer loss function ； The cross-entropy loss function, The Dice loss function; As the loss balance factor; Cross-entropy loss The calculation is as follows: For each output level First, the prediction map of this level Upsampled to the actual label The same resolution; subsequently, cross-entropy loss is calculated based on voxel-level classification accuracy. : ； In the formula, The total number of voxels; The number of categories; As an indicator function, when voxels Category The value is 1 if it is true, and 0 otherwise. To predict probabilities; Dice loss function The calculation is as follows: ； In the formula, This indicates the number of elements in the intersection of the predicted graph and the true labels; and These represent the number of elements in the predicted image and the actual label, respectively.

Images