Mamba-based 3D medical image mask autoencoding method

Abstract

The present application relates to a kind of 3D medical image mask self-encoding methods based on Mamba, to optimize the self-encoding task of processing 3D medical image, especially solve the problem of high memory requirement and computing resource consumption encountered when processing long sequence data.The Mamba network architecture introduced in the present application is used as the backbone network of the self-encoding model, to optimize the ability of the model to process long sequence data.Mamba is an advanced architecture based on state space model, which combines linear dynamic system theory and neural network concepts, effectively capturing the timing information and dynamic characteristics in the data.The core advantages of the present application include selective scanning algorithm, hardware-aware algorithm, and efficient data processing capability, especially when processing long sequence data.The present application not only reduces the burden of computing and storage, but also maintains or improves the accuracy and efficiency of image processing, especially suitable for processing 3D medical image data that requires a large amount of computation and high memory bandwidth.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of medical image processing, and particularly relates to a 3D medical image mask auto-encoding method based on Mamba. The present application uses a method based on the Mamba network architecture to optimize the auto-encoding task of 3D medical images (such as CT, MRI), overcoming the problem of heavy computing burden and high memory requirement caused by large data volume and long sequence length in the prior art. BACKGROUND

[0002] Structured 3D medical image data, such as computed tomography (CT) and magnetic resonance imaging (MRI), are indispensable tools in modern medical diagnosis and treatment. These images provide detailed views of the internal structure of the human body, which is of great significance for disease diagnosis, treatment planning and efficacy evaluation.

[0003] Although 3D medical image technology plays a core role in the medical field, its processing and analysis still face major challenges. In particular, the processing of 3D image data requires highly complex computing resources, as these data are not only voluminous but also structurally complex. Traditional 2D image processing techniques cannot be directly applied to 3D data, so special techniques need to be developed to process such images.

[0004] In recent years, deep learning-based methods, particularly those using Transformers, have been proposed for the analysis and processing of 3D medical images. Transformer models are favored for their advantages in handling sequential data, but they face significant challenges when dealing with long sequences. In the context of 3D medical images, the serialization of images (i.e., converting 3D voxel data into a sequential form for input into the model) results in extremely long data sequences, which not only leads to a dramatic increase in computing costs but also greatly increases the memory requirements.

[0005] Specifically, existing Transformer models require a large amount of memory and computing time when processing these long sequences, which is often unacceptable in practical medical environments. Furthermore, although techniques such as patching are used to reduce the size of input data, these methods often lose important spatial information, thereby affecting the quality of the final image reconstruction.

[0006] Therefore, there is a need for a new technology that can effectively process long sequence data of 3D medical images, which not only reduces the burden of computing and storage, but also maintains or improves the accuracy and efficiency of image processing. SUMMARY

[0007] The application provides a method for 3D medical image mask autoencoding based on Mamba, aiming to solve the problem of high memory requirement and computational resource consumption when using traditional deep learning models to process 3D medical images, especially when processing long sequence data.

[0008] The application introduces the Mamba network architecture as the backbone network of the autoencoding model, optimizing the model's ability to process long sequence data. Mamba is an advanced architecture based on state space model (SSM), which combines linear dynamic system theory and neural network concepts, effectively capturing the timing information and dynamic characteristics of data. The core advantages of Mamba include:

[0009] Selective scanning algorithm: Mamba can filter relevant and irrelevant information by selectively retaining or ignoring information in input data, improving the prediction accuracy and efficiency of the model.

[0010] Hardware-aware algorithm: Mamba optimizes memory management through parallel scanning, kernel fusion, and recalculation, reducing the demand for cache and memory. This hardware optimization makes Mamba particularly suitable for processing 3D medical image data that requires a large amount of computation and high memory bandwidth.

[0011] Efficient data processing: Mamba uses the theory of state space model to simplify data processing through matrix operations, significantly improving processing speed, especially when processing long sequence data.

[0012] The application provides a method for 3D medical image mask autoencoding based on Mamba, which includes the following steps:

[0013] Step 1: Collection and organization of 3D medical image data: Obtain 3D medical images including CT, MRI, etc., and organize the data for subsequent processing.

[0014] Step 2: Image preprocessing: Preprocess all collected 3D medical image data, including image registration, image resampling, image normalization, etc., to ensure that all images have consistent dimensions and pixel value ranges.

[0015] Step 3: Data augmentation: Perform various data augmentation operations on the preprocessed 3D medical image data to increase data diversity and model generalization ability. Specific data augmentation methods include:

[0016] CutMix: Randomly select regions in the image and replace them with corresponding parts of other images.

[0017] Random rotation: Rotate the image at a random angle.

[0018] Random Flip: Randomly perform horizontal or vertical flip.

[0019] Affine Transform: Perform affine transform on the image, change the size, rotation angle and offset of the image.

[0020] These enhancement operations can be controlled by setting the probability parameters to control the frequency and degree of execution.

[0021] Step 4: Construct an auto-encoding model based on the Mamba architecture, specific structure:

[0022] The core model architecture of the auto-encoding model is the Mamba model. The core building block of the Mamba model is the Mamba block (MambaBlock). By stacking multiple Mamba blocks, a deep Mamba model is constructed.

[0023] The Mamba block is the basic component unit of the Mamba model, as Figure 4 shown, a typical Mamba block is mainly composed of the following key components: input projection layer, normalization layer, selective SSM layer, output projection layer and residual connection. The Mamba block receives the input sequence x(t), which is processed in turn through the input projection layer, the normalization layer, the selective SSM layer and the output projection layer. Finally, the original input sequence x(t) is added to the output of the output projection layer through the residual connection to obtain the Mamba block output. The Mamba block output is used as the input of the next Mamba block, or as the final output of the Mamba model.

[0024] Mamba block component 1: input projection layer. The input projection layer is the first component of the Mamba block, responsible for dimension transformation of the input sequence. The input projection layer mainly includes a linear layer for linear transformation of the input sequence x(t) to project it into a higher dimensional feature space. GELU is added as an activation function after the linear layer to introduce nonlinearity. The output of the input projection layer is the projected feature, which will be used as the input of the subsequent normalization layer.

[0025] Mamba block component 2: normalization layer. The normalization layer is used to stabilize the training process and accelerate model convergence. The normalization layer uses Layer Normalization to perform normalization operation on the input features. The normalization operation normalizes the distribution of the features to obtain the normalized features as the input of the selective SSM layer.

[0026] Mamba block component 3: selective SSM layer. The selective SSM layer is the core component of the Mamba model, responsible for sequence modeling and selective information extraction. The selective SSM layer receives input from the normalization layer and processes it through the selective state transition mechanism, state update unit and output mapping unit in turn.

[0027] Further, Selective SSM Layer Subcomponent 3.1: Selective State Transition Mechanism. The Selective State Transition Mechanism enables the Mamba model to dynamically adjust state transition behavior according to input. Its function is to generate input-dependent state transition matrix A(x(t)) and input matrix B(x(t)). The specific implementation is to project the input into the parameter space through a linear layer, which is used to dynamically parameterize the state transition matrix A(x(t)) and the input matrix B(x(t)).

[0028] Further, Selective SSM Layer Subcomponent 3.2: State Update Unit. The State Update Unit is responsible for calculating the state at the current time according to the current input and the state at the last time. The state update process follows the formula: h(t) = A(x(t))h(t-1) + B(x(t))x(t). Where h(t) is the state vector at time t, x(t) is the input vector at time t, A(x(t)) and B(x(t)) are the input-dependent state transition matrix and input matrix generated by the Selective State Transition Mechanism.

[0029] Further, Selective SSM Layer Subcomponent 3.3: Output Mapping Unit. The Output Mapping Unit is responsible for mapping the state vector h(t) to the output y(t) of the SSM layer. The output mapping process follows the formula: y(t) = Ch(t). Where C is the output matrix.

[0030] Mamba Block Component 4: Output Projection Layer. The Output Projection Layer receives the output from the Selective SSM Layer and performs dimension transformation to match the requirements of subsequent modules or tasks. The Output Projection Layer is similar in structure to the Input Projection Layer, mainly containing a linear layer for linear transformation of input features (SSM output) to adjust feature dimensions. Nonlinearity is introduced after the linear layer by adding GELU as the activation function. The output of the Output Projection Layer is the output projection, which will be used for residual connection or as the final output of the Mamba block.

[0031] Mamba Block Component 5: Residual Connection. The Residual Connection adds the input of the Mamba block directly to the output of the Output Projection Layer to form the final output of the Mamba block.

[0032] The Mamba model used in the method of the present application is stacked by multiple Mamba blocks described above. By stacking Mamba blocks, the model can gradually extract deep features of 3D medical image sequences and be used in the task of auto-encoding reconstruction. The number of Mamba block stacks and the parameter configuration inside each block are adjusted and optimized according to specific application scenarios and task requirements.

[0033] Unlike Transformer models, Mamba models lack global contextual information, while the spatial contextual information of medical 3D images is extremely complex and crucial. To more effectively capture the spatial contextual information of 3D medical images and fully leverage the advantages of Mamba models in sequence modeling, this invention introduces a multi-strategy scanning serialization method for 3D images between stacked Mamba blocks. This method aims to compensate for the shortcomings of the native structure of the Mamba model in global context awareness through multi-angle and multi-directional serialization scanning, enhancing the model's ability to capture 3D spatial information and thus improving the performance of 3D medical image autoencoding.

[0034] Furthermore, the multi-strategy scan serialization method proposed in this invention, such as Figure 5 As shown, this invention aims to serialize 3D medical images from multiple angles and directions to capture their spatial context information more comprehensively. Specifically, this invention designs eight different image scanning serialization methods.

[0035] The multi-strategy scan serialization method is implemented as follows:

[0036] Step (1): Divide the 3D medical image into slices. First, divide the input 3D medical image data into a series of 2D slices along a specified axis (taking the Z-axis as an example). Each slice represents information of the 3D image at a specific level.

[0037] Step (2): Perform multi-directional image patch division within each slice. For each 2D slice, first fix the image patch size, and then use four different image patch division strategies to extract image patch sequences from different directions:

[0038] 2.1: Starting from the top left corner, X-axis priority scanning: Starting from the top left corner of the slice (x=0, y=0), the slice is first divided into image block sequences along the X-axis direction, and then moved along the Y-axis direction and continued to divide along the X-axis direction until the entire slice is covered.

[0039] 2.2: Starting from the top left corner, Y-axis priority scanning: Starting from the top left corner of the slice (x=0, y=0), the slice is first divided into image block sequences along the Y-axis direction, and then moved along the X-axis direction and continued to divide along the Y-axis direction until the entire slice is covered.

[0040] 2.3: Starting from the lower right corner, reverse X-axis priority scanning: Starting from the lower right corner of the slice, first divide the slice into image block sequences along the reverse X-axis direction, then move along the reverse Y-axis direction and continue to divide along the reverse X-axis direction until the entire slice is covered.

[0041] 2.4: Bottom-right corner start, reverse Y-axis first scan: Starting from the bottom-right corner of the slice, first divide the slice into image block sequences along the reverse Y-axis direction, then move along the reverse X-axis direction and continue the division along the reverse Y-axis direction until the entire slice is covered.

[0042] Step (3): Constructing multi-directional slice sequences. After completing the four image block division methods within a single slice, combine two different slice orders to construct the final 3D medical image sequence:

[0043] 3.1: Forward slice order: Arrange the slices in the original order from the first slice to the last slice.

[0044] 3.2: Reverse slice order: Arrange the slices in the reverse order from the last slice to the first slice.

[0045] Step (4): Combine eight scanning sequence methods. By combining the four single-slice image block division methods in Step 2 with the two slice orders in Step 3, a total of (2+2)*2=8 different 3D medical image scanning sequence methods are obtained. These eight methods sequence the 3D image from different angles and directions, forming a multi-strategy scanning sequence set.

[0046] The multi-strategy scanning sequence method is combined with the Mamba block, specifically by adding a scanning sequence rearrangement module before each Mamba block. During model initialization, each scanning sequence rearrangement module is statically assigned to one of the eight scanning sequence methods described above. This assignment remains unchanged during subsequent training and testing. The scanning sequence rearrangement module rearranges the input 3D medical image sequence according to the preset strategy.

[0047] When stacking Mamba blocks to form a Mamba model, an alternating stacking architecture is used. That is, the scanning sequence rearrangement module and the Mamba block are alternately stacked. After the data flows through the scanning sequence rearrangement module, it is rearranged according to the specific scanning order and then input into the Mamba block for feature learning. This alternating structure allows the model to observe the 3D image from different scanning angles at each processing stage, integrating multi-directional context information. The number of stacked Mamba blocks is an integer multiple of 8, corresponding to an integer multiple of the eight scanning sequence methods.

[0048] Step 5: Model training: Use data-augmented images as input and original unobstructed images as labels to train the Mamba architecture-based auto-encoding model. The training of this model includes using random occlusion as a strategy to simulate missing data, thereby improving the model's ability to handle partial missing data in actual applications.

[0049] Step 6: Image reconstruction and new sample encoding: Use the trained autoencoder model to encode and reconstruct new medical image samples, test the model's application effect in the actual medical environment, and ensure that the medical images can be reconstructed with high quality.

[0050] Step 7: Loss function and optimizer settings:

[0051] Optimizer: The model training uses the Adam optimizer for parameter optimization, and sets appropriate learning rate and iteration times to ensure model convergence.

[0052] Further, the present application provides a 3D medical mask autoencoding system, including a trained and verified mask autoencoding network.

[0053] Further, the present application also provides a computer readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to execute the method.

[0054] Further, the present application also provides a computing device including a memory and a processor, the memory having executable code stored therein, and the processor executing the executable code to implement the method.

[0055] The beneficial effects of the present application are:

[0056] Improve processing efficiency and reduce cost: By using the Mamba structure autoencoder instead of the traditional Transformer structure, the present application significantly improves the efficiency of 3D medical image processing. Since the Mamba structure optimizes the processing of long sequence data, it reduces the demand for video memory and computing resources, thereby also reducing the cost of data processing.

[0057] Improve image reconstruction quality: The present application uses MambaBlocks in the autoencoder, so that the model can more effectively process and reconstruct 3D medical images. Especially in the process of reconstructing occluded images, the Mamba structure can more accurately restore the details of the image through its efficient data processing capability, improving the quality of the reconstructed image.

[0058] Enhance the generalization ability of the model: The use of extensive data augmentation techniques (such as CutMix, random rotation, random flip, affine transformation, etc.) makes the trained model have better generalization ability. These techniques simulate various actual medical image conditions, so that the model can work more stably in actual application and has good adaptability to various medical image scenarios.

[0059] Adapt to large-scale data processing: The autoencoder of Mamba structure is particularly suitable for processing large-scale medical image data sets. Due to its hardware optimization and parallel processing capability, the present application can maintain high efficiency when processing a large amount of data, which is suitable for the large data demand commonly seen in modern medical image processing. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 is an image autoencoding flowchart.

[0061] Figure 2 is a model training flowchart.

[0062] Figure 3 is an architecture diagram of Mamba encoder and decoder.

[0063] Figure 4 is an architecture diagram of a Mamba block.

[0064] Figure 5 is a schematic diagram of 3D medical image multi-strategy scanning serialization. DETAILED DESCRIPTION

[0065] The present embodiment is a possible implementation based on Python. The technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all.

[0066] As shown in Figures 1-3 , the present application is a 3D medical image autoencoding method based on Mamba structure, wherein the steps include:

[0067] S1: Collection and arrangement of 3D medical images: Obtain 3D medical images including CT, MRI, etc. The recommended collection resolution is at least 512x512x512 pixels to ensure that sufficient details are captured to provide high-quality input data for subsequent processing.

[0068] S2: Image preprocessing: Preprocess all collected 3D medical images, including image registration, image resampling, image normalization, etc. The recommended resampling target resolution is 1mmx1mmx1mm, and the pixel intensity should be standardized to the range of 0-1.

[0069] S3: Data augmentation: Perform various data augmentation operations on the preprocessed images to increase the diversity of the data and the generalization ability of the model. Specific data augmentation methods include:

[0070] CutMix: Randomly select an area proportion of 20% of the total image area and replace it with the corresponding part of another image.

[0071] Random rotation: rotation angle is [-10°, +10°].

[0072] Random flip: with 50% probability, either horizontal or vertical flip.

[0073] Affine transformation: transformation parameters are set to adjust size within [-10%, +10%], rotation within [-10°, +10°], and image translation up to 10%.

[0074] S4: Model initialization: Before entering the model training (S5), the constructed 3D auto-encoding model based on Mamba structure needs to be initialized. The initialization process mainly includes setting model architecture parameters and initializing the learnable parameters of each component. The number of stacked Mamba blocks is set to 24 layers (3*8), and the scan sequence rearrangement module before each Mamba block is randomly assigned one strategy, but it is ensured that each strategy appears three times in the overall model. The trainable parameters in the model are initialized with Xavier to ensure the stability of the training. This initialization method is one of the possible implementations of the invention, not all initialization methods.

[0075] S5: Model training: using data-augmented images as input and original unoccluded images as label, train the auto-encoding model based on Mamba structure. The detailed training method of the flow chart is shown in Figure 2 During training, it is recommended to set the batch size to 32, the learning rate to 0.001, and the number of iterations to at least 500 epochs to ensure sufficient learning.

[0076] S6: Image reconstruction and new sample encoding: use the trained auto-encoding model to encode and reconstruct new medical image samples. The key to this step is to ensure that the model can correctly perform in the actual medical environment, so it is recommended to verify on multiple independent test sets.

[0077] S7: Loss function and optimizer settings:

[0078] Optimizer: Adam optimizer is used for parameter optimization during model training, with an initial learning rate of 0.001 and default beta parameters (β1=0.9, β2=0.999). To optimize the training process, a cosine decay learning rate scheduling mechanism with warmup is adopted. The specific settings are as follows: the learning rate is linearly increased to the set maximum value within the first 10 epochs of training, and then gradually decreases according to the cosine function until the end of training.

[0079] System Deployment and Validation: After completing the training and testing of the model, the system is deployed. During deployment, it is necessary to ensure that all hardware and software configurations can meet the processing needs, including sufficient computing power and storage space. System validation should be performed through comparative analysis with existing diagnostic processes to ensure that the performance of the new system is at least comparable to the existing system.

[0080] Subsequent Maintenance and Upgrade: After the system is deployed, regular maintenance and upgrades are carried out to meet new medical image processing needs and technological development. Maintenance activities include monitoring system performance, regularly checking and replacing faulty hardware, and updating software to fix known problems or improve performance.

Claims

1. A method of Mamba-based 3D medical image mask autoencoding, characterized in that, The method comprises the following steps: Step 1: 3D medical image data acquisition and arrangement: acquire 3D medical images including CT and MRI types, and arrange the data; Step 2: Image preprocessing: preprocess all collected 3D medical image data, including image registration, image resampling, image normalization, to ensure that all images have consistent size and pixel value range; Step 3: Data augmentation: perform various data augmentation operations on the preprocessed 3D medical image data to increase data diversity and model generalization ability; Step 4: Constructing an auto-encoding model based on the Mamba architecture, the specific structure is: The core model architecture of the auto-encoding model is the Mamba model, and the core building block of the Mamba model is the Mamba block. By stacking multiple Mamba blocks, a deep Mamba model is constructed. A multi-strategy scan serialization method for 3D images is introduced between the Mamba blocks, that is, the multi-strategy scan serialization method is combined with the Mamba block. Specifically, a scan serialization rearrangement module is added before each Mamba block. During model initialization, each scan sequence rearrangement module is statically assigned to one of the eight scan serialization methods. This assignment remains unchanged during subsequent training and testing. The scan sequence rearrangement module rearranges the input 3D medical image sequence according to the preset strategy. When stacking Mamba blocks to form a Mamba model, an alternating stacking architecture is adopted. That is, the scan sequence rearrangement module and the Mamba block are alternately stacked. After the data flows through the scan sequence rearrangement module, it is rearranged in a specific scan order and then input into the Mamba block for feature learning. This alternating structure allows the model to observe 3D images from different scan angles at each processing stage, integrating multi-directional context information. The number of stacked Mamba blocks is an integer multiple of 8, corresponding to an integer multiple of the eight scan sequence methods. Step 5: Model training: use data-enhanced images as input and original unoccluded images as labels to train the auto-encoding model based on the Mamba architecture. The training of this model includes using random occlusion as a strategy to simulate missing data, thereby improving the model's ability to handle partial missing data in actual applications. Step 6: Image reconstruction and new sample encoding: use the trained auto-encoding model to encode and reconstruct new medical image samples to test the model's application effect in actual medical environments and ensure that medical images can be reconstructed with high quality. Step 7: Loss function and optimizer setting.

2. The method of Mamba-based 3D medical image mask autoencoding according to claim 1, characterized in that, The Mamba block is composed of the following key components: input projection layer, normalization layer, selective SSM layer, output projection layer, and residual connection. The Mamba block receives input sequence x(t) and processes it through the input projection layer, normalization layer, selective SSM layer, and output projection layer in turn. Finally, the original input sequence x(t) is added to the output of the output projection layer through the residual connection to obtain the Mamba block output. The Mamba block output serves as the input of the next Mamba block or as the final output of the Mamba model.

3. The method of Mamba-based 3D medical image mask autoencoding according to claim 2, characterized in that, The selective SSM layer is the core component of the Mamba model, responsible for sequence modeling and information selective extraction. The selective SSM layer receives input from the normalization layer and processes it through a selective state transition mechanism, a state update unit, and an output mapping unit in turn.

4. The method of claim 3, wherein, The selective state transition mechanism enables the Mamba model to dynamically adjust state transition behavior according to input. Its function is to generate input-dependent state transition matrix A(x(t)) and input matrix B(x(t)); the specific implementation is to project the input into the parameter space through a linear layer, which is used to dynamically parameterize the state transition matrix A(x(t)) and the input matrix B(x(t)).

5. The method of claim 3, wherein the Mamba-based 3D medical image mask autoencoding is characterized by, The state update unit is responsible for calculating the current state according to the current input and the state at the previous time. The state update process follows the formula: h(t) = A(x(t))h(t-1) + B(x(t))x(t); where h(t) is the state vector at time t, x(t) is the input vector at time t, and A(x(t)) and B(x(t)) are the input-dependent state transition matrix and input matrix generated by the selective state transition mechanism.

6. The method of claim 3, wherein the Mamba-based 3D medical image mask autoencoding is characterized by, The output mapping unit is responsible for mapping the state vector h(t) to the output y(t) of the SSM layer; the output mapping process follows the formula: y(t) = Ch(t); where C is the output matrix.

7. The method of claim 1, wherein the Mamba-based 3D medical image mask autoencoding is characterized by, The multi-strategy scanning sequence method is implemented as follows: Step (1): Divide the 3D medical image slices; first, divide the input 3D medical image data into a series of 2D slices along the specified axis; each slice represents the information of the 3D image at a specific level; Step (2): Perform multi-directional image block division within each slice; for each 2D slice, first fix the image block size and use four different image block division strategies to extract image block sequences from different directions; Step (3): Construct multi-directional slice sequences; after completing the four image block division strategies within a single slice, combine two different slice orders to construct the final 3D medical image sequence; Step (4): Combine eight scanning sequence methods; by combining the four single-slice image block division methods in step (2) with the two slice orders in step 3, a total of (2+2)*2=8 different 3D medical image scanning sequence methods are obtained; these eight methods sequence the 3D image from different angles and directions, forming a multi-strategy scanning sequence set.

8. The method of Mamba-based 3D medical image mask autoencoding according to claim 7, characterized in that, Step (2) extracts image block sequences from different directions as follows: 2.1: Left upper corner starting point, X-axis priority scanning: taking the left upper corner of the slice as the starting point, first divide the slice into image block sequences along the X-axis direction, then move along the Y-axis direction and continue the X-axis direction division until the entire slice is covered; 2.2: Left upper corner starting point, Y-axis priority scanning: taking the left upper corner of the slice as the starting point, first divide the slice into image block sequences along the Y-axis direction, then move along the X-axis direction and continue the Y-axis direction division until the entire slice is covered; 2.3: right-bottom corner start, reverse X-axis priority scanning: taking the right-bottom corner of the slice as the starting point, first dividing the slice into a sequence of image blocks along the reverse X-axis direction, then moving along the reverse Y-axis direction and continuing the division along the reverse X-axis direction until the entire slice is covered; 2.4: right-bottom corner start, reverse Y-axis priority scanning: taking the right-bottom corner of the slice as the starting point, first dividing the slice into a sequence of image blocks along the reverse Y-axis direction, then moving along the reverse X-axis direction and continuing the division along the reverse Y-axis direction until the entire slice is covered.

9. The method of Mamba-based 3D medical image mask autoencoding according to claim 7 or 8, characterized in that, Step (3) constructs the final 3D medical image sequence, which is as follows: 3.1: forward slice order: arranging the slices in the original order from the first slice to the last slice; 3.2: reverse slice order: arranging the slices in the reverse order from the last slice to the first slice.

10. A Mamba-based 3D medical image mask auto-encoding system, characterized in that, The Mamba architecture-based self-encoding model trained and verified according to claim 1.

Images