Paranasal sinus bone structure medical image segmentation system based on Mamb-UNet

Abstract

The invention provides a paranasal sinus bone structure medical image segmentation system based on Mamb-UNet, and relates to the technical field of medical image processing.The paranasal sinus bone structure medical image segmentation system comprises the steps that firstly, a low-resolution global abstract graph is extracted, original features are injected, and a local extractor is endowed with context sensing capacity; and the bone slices with similar forms and different meanings can be distinguished according to anatomical positions, so that false alarms are remarkably reduced. The invention aims to ensure that the boundary of the segmented bone structure is clear by introducing a gating type bidirectional fusion mechanism and dynamically arbitrating the weights of forward and backward information streams by utilizing a learnable gating unit. According to the method, a hierarchical parallel Mama (H-Mama) architecture is designed, and parallel multi-scale feature extraction of features is realized in the same feature level. By operating high-frequency, intermediate-frequency and low-frequency branches in parallel, the model can capture fine bone wall details and macroanatomical layout simultaneously without compromise so as to adapt to violent size changes.

Description

A Medical Image Segmentation System for Paranasal Sinus Bone Structure Based on Mamba-UNet Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a medical image segmentation system for paranasal sinus bone structure based on Mamba-UNet. Background Technology

[0002] Currently, bone structure segmentation techniques in this field are mainly divided into the following three categories: (1) Traditional manual segmentation method: In current clinical practice, bone structure segmentation mainly relies on radiologists or surgeons to manually delineate on the imaging workstation. This method is not only time-consuming and laborious, but the segmentation work of a single case often takes several hours to complete. In addition, manual segmentation is highly dependent on the doctor's personal experience, and there are significant subjective differences in the segmentation results between different doctors and even between the same doctor at different times.

[0003] (2) Existing deep learning automatic segmentation methods: Although deep learning has made great progress in the field of medical image segmentation, existing bone segmentation research mainly focuses on bones with relatively simple structures and large volumes, such as long bones (e.g., femur), ribs, or large skull bones (e.g., frontal bone, parietal bone). Research on automatic segmentation of the paranasal sinuses, which are highly dense, delicate, and irregular skeletal structures (OSPS), is very scarce.

[0004] (3) Hybrid architecture based on state space model (Mamba): In order to solve the problems of limited receptive field of traditional convolutional neural networks (CNN) and high computational cost of Vision Transformer (ViT), state space models (SSMs), especially the Mamba architecture, have been introduced in recent years. With its linear computational complexity and excellent long sequence modeling advantages, the Mamba architecture has shown the potential to surpass ViT in capturing the global dependencies of high-resolution medical images.

[0005] While Mamba-UNet performs well on general tasks, it suffers from the following core technical limitations when processing paranasal sinus bone structures (OSPS): 1. "Context blindness" in local feature extraction: Existing architectures typically employ a sequential processing paradigm of local features first, then global features. This means that local extractors lack prior guidance on macroscopic anatomical locations when processing high-resolution details (such as tiny bone fragments).

[0006] 2. "Signal contamination" in bidirectional sequence modeling: To address the orientation sensitivity of images, existing methods typically perform bidirectional scanning (forward + backward) and then fuse the images using a simple "element-by-element addition". Due to memory decay in SSM (sensitive to information at the end of the sequence and blurred information at the beginning), simple addition causes the high signal-to-noise ratio features at both ends of the sequence (from the end of the scan) to be contaminated by the low signal-to-noise ratio features in the other direction (from the beginning of the scan).

[0007] 3. "Scale mismatch" in single-scale modeling: The target scale range in the paranasal sinus region is extremely wide (from the massive maxillary sinus to the tiny ethmoid air cells). Existing Mamba modules typically use a fixed-scale processing window. An overly large window smooths out crucial micro-bone walls, while an overly small window fails to capture extensive anatomical variations. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the present invention aims to propose a medical image segmentation system for paranasal sinus bone structures based on Mamba-UNet, comprising a data acquisition module and an image segmentation module. The data acquisition module is used to acquire medical images of paranasal sinus bone structures and send these images to the image segmentation module. The image segmentation module is used to receive the medical images of paranasal sinus bone structures, input them into a hybrid neural network model H-IHF-UNet, and obtain segmented images. The hybrid neural network model includes an encoder and a decoder. The encoder includes multiple hierarchical fusion blocks, and the decoder contains multiple CNN fusion modules.

[0009] Optionally, the medical images of the paranasal sinus bone structure are input into the hybrid neural network model H-IHF-UNet to obtain segmented images. This includes: inputting the medical images of the paranasal sinus bone structure into the encoder of the hybrid neural network model H-IHF-UNet to obtain encoded images; specifically, inputting the medical images of the paranasal sinus bone structure into a layered fusion block to obtain a first layered image, downsampling the first layered image to obtain a downsampled first layered image, inputting the downsampled first layered image into the layered fusion block to obtain a second layered image, downsampling the second layered image to obtain a downsampled second layered image, inputting the downsampled second layered image into the layered fusion block to obtain a third layered image, downsampling the third layered image to obtain a downsampled third layered image, and inputting the downsampled third layered image into the layered fusion block. The process involves: 1) obtaining a fourth layer image; 2) downsampling the fourth layer image to obtain an encoded image; 3) passing the encoded image through the CNN fusion module of the decoder to obtain a segmented image. Specifically, 4) upsampling the encoded image to obtain the decoder's input image; 5) concatenating the decoder's input image with the third layer image; 6) convolving the concatenated image to obtain a first fused image; 7) upsampling the first fused image to obtain an upsampled first fused image; 8) concatenating the upsampled first fused image with the second layer image; 9) convolving the concatenated image to obtain a second fused image; 10) upsampling the second fused image with the first layer image; 11) convolving the concatenated image to obtain a third fused image; and 12) passing the third fused image through a segmentation layer to obtain a segmented image.

[0010] Optionally, medical images of the paranasal sinus bone structure are input into a layered fusion block to obtain a first layered image, including: medical images of the paranasal sinus bone structure. After the average pooling layer, the summary image is obtained. ; The summary graph is processed using a Mamba module that includes a gated bidirectional fusion mechanism. The process is performed to obtain a summary diagram after global modeling. The summary diagram after global modeling processing The global context feature map is obtained by upsampling using bilinear interpolation. , global context feature map Adding the medical images of the paranasal sinus bone structure to obtain a global information feature map. ; global information feature map The H-Mamba module is used to extract local details in a refined manner, resulting in the first layer image.

[0011] Optionally, the summary graph can be processed using a Mamba module that includes a gated bidirectional fusion mechanism. The process is performed to obtain a summary diagram after global modeling. This includes: abstract diagrams Forward features were obtained by performing forward Mamba scans and backward Mamba scans respectively. and backward features ; forward features and backward features The feature maps are concatenated along the channel dimension, and then passed through a fully connected layer and a sigmoid activation function to obtain a gated weight map. ; through gating weight graph Forward features and backward features Perform element-wise weighted summation to obtain the summary graph after global modeling. .

[0012] Optionally, through a gating weight graph Forward features and backward features Perform element-wise weighted summation to obtain the summary graph after global modeling. Specifically, this is achieved through the following formula: .

[0013] Optionally, the global information feature map The input to the H-Mamba module is used for refined extraction of local details to obtain the first layer image, including: extracting global information feature maps. After passing through high-frequency, mid-frequency, and low-frequency branches respectively, high-frequency, mid-frequency, and low-frequency feature maps are obtained. The mid-frequency and low-frequency feature maps are upsampled, and then concatenated along the channel dimension to obtain a first concatenated feature map. This first concatenated feature map is then processed... After convolutional layers and global information feature maps Add them together to obtain the first layer image.

[0014] Optionally, the global information feature map After passing through high-frequency, mid-frequency, and low-frequency branches respectively, high-frequency feature maps, mid-frequency feature maps, and low-frequency feature maps are obtained, including: global information feature maps. After high-frequency branching, a high-frequency feature map is obtained; specifically, this is achieved through the global information feature map. Divide the size by the first size to get the first number of blocks, and then use the first number of blocks to create a feature map of the global information. The data is divided into multiple high-frequency sub-feature maps. These high-frequency sub-feature maps are then processed through a Mamba module containing a gated bidirectional fusion mechanism to obtain the high-frequency feature map. The global information feature map is then processed. After the intermediate frequency branch, the intermediate frequency feature map is obtained; specifically, the global information feature map... Downsampling by a factor of 4 yields a downsampled feature map. Dividing this downsampled feature map by the second size gives the second number of blocks. The downsampled feature map is then divided according to this second number of blocks, resulting in multiple intermediate frequency (IF) sub-feature maps. These IF sub-feature maps are then processed through a Mamba module containing a gated bidirectional fusion mechanism to obtain the final IF feature map. The global information feature map is then... After the low-frequency branch, a low-frequency feature map is obtained; specifically, this is the global information feature map. Downsampling by 16 times yields a downsampled feature map. The downsampled feature map is then divided by the third size to obtain the third number of blocks. The downsampled feature map is then divided according to the third number of blocks to obtain multiple low-frequency sub-feature maps. These multiple low-frequency sub-feature maps are then processed through a Mamba module containing a gated bidirectional fusion mechanism to obtain a low-frequency feature map.

[0015] The beneficial effects of adopting the above technical solution are as follows: 1. Overcoming the context-blindness of local feature extraction and reducing the clinical false alarm rate: Addressing the problem that existing technologies employ a "local first, global later" sequential paradigm, resulting in a lack of macroscopic vision in the local extractor, the purpose of this invention is to establish a data flow mechanism that prioritizes the global and refines the local features by proposing an Interactive Hierarchical Fusion (IHF) strategy. By first extracting a low-resolution global summary map and injecting the original features, the local extractor is endowed with context-aware capabilities, enabling it to distinguish morphologically similar but meaningful bone fragments (such as differentiating normal bone septa from pathological osteophytes) based on anatomical location, thereby significantly reducing false alarms.

[0016] 2. Eliminating signal contamination in bidirectional sequence modeling to achieve millimeter-level boundary accuracy: Addressing the problem in existing technologies where simple superposition of bidirectional scanning features leads to noise contamination of high signal-to-noise ratio signals, resulting in blurred bone wall edges, this invention aims to introduce a gated bidirectional fusion (GBF) mechanism. This mechanism utilizes learnable gating units to dynamically arbitrate the weights of forward and backward information flows. This ensures clear boundaries of the segmented bone structures (especially the paper-thin orbital plate), meeting the stringent millimeter-level accuracy requirements of surgical "no-entry zones."

[0017] 3. Addressing scale mismatch in complex anatomical structures and enhancing robustness of multi-scale targets: Addressing the issue of structural compromises caused by existing models using only fixed windows in the paranasal sinus region, which has a vast target scale span (from the enormous maxillary sinus cavity to the tiny ethmoid air chamber), this invention aims to achieve parallel multi-scale feature extraction within the same feature level by designing a hierarchical parallel Mamba (H-Mamba) architecture. By running high-frequency (detail), mid-frequency (structure), and low-frequency (macro) branches in parallel, the model can simultaneously and uncompromisingly capture both fine bone wall details and macroscopic anatomical layout, thus adapting to dramatic size changes.

[0018] 4. Provide high-precision automated tools to assist in clinical diagnosis and treatment: Ultimately, this invention aims to fill the technological gap in the automatic segmentation of the complex, delicate and irregular bone structures around the paranasal sinuses, and to provide an automated method that can quickly and accurately delineate key anatomical structures without human intervention, so as to assist doctors in reducing surgical risks and avoiding serious complications such as blindness and cerebrospinal fluid leakage. Attached Figure Description

[0019] Figure 1 is a schematic diagram of a medical image segmentation system for paranasal sinus bone structure based on Mamba-UNet in an embodiment of the present invention; Figure 2 is a schematic diagram of the hybrid neural network model H-IHF-UNet in an embodiment of the present invention; Figure 3 is a schematic diagram of the layered fusion block IHF Block in an embodiment of the present invention; Figure 4 is a schematic diagram of the Mamba module containing a gated bidirectional fusion (GBF) mechanism in an embodiment of the present invention; Figure 5 is a schematic diagram of the H-Mamba module in an embodiment of the present invention. Detailed Implementation

[0020] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0021] To address the problems existing in the prior art, this invention provides a medical image segmentation system for paranasal sinus bone structure based on Mamba-UNet, applied to preoperative planning and intraoperative navigation systems for endoscopic sinus surgery and skull base surgery. Referring to Figure 1, it includes a data acquisition module and an image segmentation module; the data acquisition module is used to acquire medical images of paranasal sinus bone structure and send them to the image segmentation module; the image segmentation module is used to receive the medical images of paranasal sinus bone structure and input them into a hybrid neural network model H-IHF-UNet to obtain segmented images; the hybrid neural network model includes an encoder and a decoder, the encoder including multiple hierarchical fusion blocks, and the decoder containing multiple CNN fusion modules; specifically, the medical images of paranasal sinus bone structure are input into the encoder of the hybrid neural network model H-IHF-UNet to obtain encoded images; specifically, referring to Figure 2, the medical images of paranasal sinus bone structure are input into the hierarchical fusion blocks to obtain the first layer image; referring to Figure 3, specifically including: to reduce computational complexity, the medical images of paranasal sinus bone structure... After an average pooling layer, the feature map is compressed into a low-resolution summary map. Obtain the summary diagram Where B represents the batch size, C represents the number of channels, H represents the height, and W represents the width; to maintain computational efficiency, a pooling factor is used. For example, for input resolution ,set up To ensure the resolution of the summary figure The size is relatively small, thus generating short sequences.

[0022] Based on this, medical imaging of the paranasal sinus bone structure After the average pooling layer, it can be expressed by the following formula: ;in, Indicates the average pooling layer. This indicates the height after pooling. This represents the width after pooling; the summary graph is processed using a Mamba module that includes a gated bidirectional fusion (GBF) mechanism. The process is performed to obtain a summary diagram after global modeling. Thus, the model can capture the absolute global dependencies of an image with extremely low resource consumption. In order to solve the "signal pollution" problem caused by memory decay during the scanning process of bidirectional SSM (i.e., clear features at the end of the sequence are interfered with by blurry features at the beginning of the other direction), this invention introduces a learnable position-aware gating unit.

[0023] Specifically, referring to Figure 4, the abstract diagram... Forward features were obtained by performing forward Mamba (SSM) scans and backward Mamba (SSM) scans respectively. and backward features Among them, forward features Rich information accumulates at the end of the sequence, but less at the beginning. Backward features. The sequence accumulates a wealth of information at the beginning, but less at the end.

[0024] forward features and backward features Concatenate along the channel dimension, then pass the concatenated feature maps through a fully connected layer and a sigmoid activation function to obtain a set of features with values ​​ranging from [value range missing]. Gating weight graph between Specifically, this is achieved through the following formula: ;in, Indicates the gating weight, This indicates a splicing operation; the network automatically learns to recognize patterns with "high signal-to-noise ratio" characteristics through training. The higher the value, the clearer the forward features at that location, according to the model; conversely, the clearer the backward features.

[0025] Through the gating weight graph Forward features and backward features Perform element-wise weighted summation to obtain the summary graph after global modeling. Specifically, this is achieved through the following formula: It should be noted that the Mamba module with gated bidirectional fusion (GBF) mechanism mentioned in this invention has the same processing operation as described above, only the input and output are different.

[0026] Summary diagram after global modeling Upsample back to the original resolution using bilinear interpolation. The global context feature map is obtained. Subsequently, the global context feature map... Adding the medical images of the paranasal sinus bone structure to obtain a global information feature map. Specifically, it is expressed by the following formula: ; ; what was obtained at this time Each pixel in the feature map explicitly carries macroscopic anatomical coordinate information (e.g., "this pixel is located at the center of the sinus" or "at the edge of the orbit"), thereby guiding subsequent local extractors to perform targeted processing.

[0027] Global information feature map The H-Mamba module is used to refine the local details and obtain the first layer image. This design achieves the acquisition of the global field of view first and then the extraction of local features. In order to solve the problem of target scale mismatch in the paranasal sinus region (i.e., the simultaneous existence of a huge maxillary sinus and a small ethmoid airway), this invention designs a parallel multi-scale feature extraction architecture in the local refinement stage of IHF.

[0028] This module contains within the same feature level One (in this embodiment, take) The parallel processing branches correspond to different resolution features and receptive fields.

[0029] Specifically, referring to Figure 5, the global information feature map After passing through high-frequency, mid-frequency, and low-frequency branches respectively, high-frequency feature maps, mid-frequency feature maps, and low-frequency feature maps are obtained; specifically, the global information feature map... After high-frequency branching ( This yields high-frequency feature maps; specifically, through global information feature maps. Divide the size by the first size to get the first number of blocks, and then use the first number of blocks to create a feature map of the global information. The data is divided into multiple high-frequency sub-feature maps. These high-frequency sub-feature maps are then processed through a Mamba module containing a gated bidirectional fusion (GBF) mechanism to obtain the high-frequency feature map. The global information feature map is then processed. After mid-frequency branch This yields the mid-frequency feature map; specifically, the global information feature map. Downsampling by a factor of 4 yields a downsampled feature map. Dividing this downsampled feature map by the second size gives the second number of blocks. The downsampled feature map is then divided according to this second number of blocks, resulting in multiple intermediate frequency (IF) sub-feature maps. These IF sub-feature maps are then processed through a Mamba module containing a gated bidirectional fusion (GBF) mechanism to obtain the final IF feature map. The global information feature map is then... After low-frequency branch ( (Low), to obtain the low-frequency feature map, specifically, the global information feature map. Downsampling by 16 times yields the downsampled feature map. Dividing the downsampled feature map by the third size gives the third block number. The downsampled feature map is then divided according to the third block number to obtain multiple low-frequency sub-feature maps. These multiple low-frequency sub-feature maps are then processed through a Mamba module containing a gated bidirectional fusion (GBF) mechanism to obtain the low-frequency feature map. The high-frequency branch focuses on capturing fine local textures, microvessels, and bone edges (high-frequency information), the mid-frequency branch focuses on capturing the internal structural homogeneity and texture consistency of organs (mid-frequency information), and the low-frequency branch focuses on capturing macroscopic anatomical layout, organ shape constraints, and relative positional relationships (low-frequency information).

[0030] Upsample the mid-frequency and low-frequency feature maps, then concatenate the upsampled mid-frequency, low-frequency, and high-frequency feature maps along the channel dimension to obtain the first concatenated feature map. This first concatenated feature map is then processed... After convolutional layers and global information feature maps The summation yields the first layer image, which is represented by the following formula: ; ;in, Represents high-frequency feature maps, This represents the mid-frequency feature map after upsampling. This represents the low-frequency feature map after upsampling. This represents the first concatenated feature map. This represents the first layer image; it should be noted that the processing operations in subsequent layer fusion blocks are the same as those described above, only the inputs and outputs are different.

[0031] The first layer image is downsampled to obtain the downsampled first layer image. This downsampled first layer image is then input into the layer fusion block to obtain the second layer image. The second layer image is then downsampled to obtain the downsampled second layer image. This downsampled second layer image is then input into the layer fusion block to obtain the third layer image. The third layer image is then downsampled to obtain the downsampled third layer image. This downsampled third layer image is then input into the layer fusion block to obtain the fourth layer image. The fourth layer image is then downsampled to obtain the encoded image. The encoded image is then passed through the CNN fusion module of the decoder to obtain the segmented image. Specifically, the encoded image is upsampled (through transposed convolution) to obtain the input image for the decoder. The decoder... The input image and the third layer image are concatenated. The concatenated image is then convolved to obtain a first fused image. The first fused image is upsampled (through transposed convolution) to obtain an upsampled first fused image. The upsampled first fused image and the second layer image are concatenated. The concatenated image is then convolved to obtain a second fused image. The second fused image is then upsampled (through transposed convolution) to obtain an upsampled second fused image. The upsampled second fused image and the first layer image are concatenated. The concatenated image is then convolved to obtain a third fused image. The third fused image is then passed through a segmentation layer (Finalprojection) to obtain a segmented image. This invention achieves parallel multi-scale feature extraction. This convolutional layer can adaptively adjust the weights according to the specific anatomical location (e.g., increasing the weights of high-frequency branches when segmenting small air chambers), thereby achieving robust segmentation of multi-scale targets.

[0032] This invention makes substantial improvements in network architecture design, feature fusion mechanism, and multi-scale processing strategy. Its core invention points are as follows: 1. Pioneering Global Priority, Local Refinement Information Flow Architecture (IHF) The shortcomings of existing technologies: Traditional U-Net or Mamba-UNet architectures generally follow a serial data flow of "first local extraction (high resolution), then global integration (low resolution)". This design causes shallow local feature extractors to be in a "blind" state when processing small image patches, lacking guidance from macroscopic anatomical locations, and easily generating false alarms.

[0033] The improvement of this invention lies in its reconstruction of the data flow within the encoder through an interactive hierarchical fusion (IHF) module. It innovatively proposes a processing path of summary generation -> global modeling -> local refinement.

[0034] Technical Principles and Effects: Specifically, the model first compresses high-resolution features into low-resolution "summaries" to quickly capture global dependencies and generate a global contextual feature map; then, this feature map is pre-injected into the original features. This design enables local feature extraction to perform targeted texture enhancement based on the macroscopic anatomical coordinates of pixels (such as whether they are located at the center of the sinus or the edge of the eye socket), fundamentally solving the "contextual blindness" problem of local extraction.

[0035] 2. Addressing the shortcomings of existing gated bidirectional fusion (GBF) mechanisms: Existing bidirectional SSM (state-space model) typically employs simple element-by-element addition when fusing forward and backward scan features. Due to the memory decay characteristic of SSM (clear memory at the end of the sequence, blurred memory at the beginning), this simple averaging operation leads to the clear signal being contaminated by the blurred signal (noise) in the other direction, directly causing blurring of the bone edges.

[0036] Improvements of this invention: This invention designs a position-aware gated unit to replace the traditional additive fusion.

[0037] Technical Principle and Effects: This mechanism evaluates the sharpness (signal-to-noise ratio) of the forward and backward feature flows in real time through a lightweight network. At any pixel location in the image, the model adaptively arbitrates and accepts the feature from the direction with the most information and the highest confidence. This mechanism effectively eliminates "signal contamination," ensuring that the segmented bone wall boundaries (especially the paper-thin orbital plate) maintain extremely high sharpness and continuity.

[0038] 3. Constructing a multi-scale parallel feature extraction architecture within a single layer (H-Mamba) Limitations of existing technologies: Existing Mamba modules typically rely on a single, fixed window size for modeling. When dealing with anatomical targets spanning vast areas of the paranasal sinus region (from the enormous maxillary sinus to the tiny ethmoid anterior chamber), a single scale approach results in large windows smoothing out details, while small windows lose sight of the overall structure.

[0039] The improvement of this invention is that it constructs a hierarchical parallel Mamba (H-Mamba) module within the same feature level of the network.

[0040] Technical Principles and Effects: This module operates three branches with different receptive fields in parallel, achieving feature resolution decoupling: High-frequency branch (large partition): focuses on capturing edges and fine textures; Mid-frequency branch (medium partition): focuses on maintaining the consistency of the internal structure of organs; High-frequency branch (small partition): focuses on constraining macroscopic anatomical shapes. Finally, the three are intelligently fused through learnable convolutional layers. This design breaks the limitations of fixed windows, enabling the model to simultaneously and uncompromisingly adapt to the segmentation needs of both extremely large and extremely small targets, significantly improving the robustness of the algorithm.

[0041] The H-IHF-UNet method proposed in this invention demonstrates significant technical advantages and clinical application value in the segmentation task of paranasal sinus bone structures (OSPS). Compared with existing mainstream medical image segmentation models, this invention achieves substantial progress in the following three aspects: 1. Comprehensive improvement of quantitative indicators (higher segmentation accuracy). Experimental results show that this invention significantly improves segmentation accuracy by introducing global context injection and parallel multi-scale feature extraction. The comparative test data with other medical segmentation models under the same test set conditions are as follows: Table 1 Comparison of Models and Experimental Indicators

[0042] The DICE coefficient (similarity index) of this invention (H-IHF-UNet) reached 94.03, significantly outperforming the traditional nnUNet ResNet (91.74) and existing Mamba-type models U-MAMBA Bot (92.12) and U-MAMBA Enc (92.19). This indicates that the segmentation mask generated by this invention has the highest overlap with the gold standard (Ground Truth).

[0043] Other key metrics: In terms of metrics measuring boundary fit (such as HL / Hausdorff Loss related metrics), the present invention achieved a score of 93.22, which is also superior to nnUNet (92.42) and U-MAMBA Bot (92.21). This demonstrates that the present invention has a lower error rate when dealing with complex geometries and can more realistically reproduce skeletal morphology.

[0044] 2. Extremely high fidelity of edge details (solving signal contamination) Thanks to the unique gated bidirectional fusion (GBF) mechanism of this invention, the model successfully eliminates the signal contamination problem in bidirectional SSM scanning: fine structures are clearly visible: visual comparison results (see attached figure in the specification) show that for small anatomical structures that are easily lost or blurred in existing models, such as the honeycomb-like bone wall of the ethmoid sinus and the paper-thin orbital plate, this invention can segment continuous edges without breakage or artifacts.

[0045] Precise boundary definition: Overlay comparison with Ground Truth shows that the segmentation boundaries of this invention closely match the actual anatomical boundaries. This directly proves that the GBF mechanism can adaptively preserve high signal-to-noise ratio features, effectively solving the edge blurring problem caused by SSM memory decay.

[0046] 3. Superior Multi-Scale Target Adaptability (Solving Scale Mismatch): This invention achieves parallel multi-scale feature extraction through a hierarchical parallel Mamba (H-Mamba) architecture, significantly improving the model's robustness to anatomical targets of different sizes: It balances large and small targets: Experiments show that the model can completely segment the enormous maxillary sinus cavity while accurately capturing the tiny ethmoid air chambers and narrow frontal recesses. This overcomes the shortcomings of existing single-scale models, which lose detail in large windows and lose global accuracy in small windows.

[0047] Reduced anatomical misjudgment: Because the IHF module provides a global anatomical view, the model can effectively distinguish structures with similar textures but different locations (such as accurately distinguishing bone septa from bone hyperplasia), which greatly reduces the false positive rate and ensures the anatomical rationality of the segmentation results.

[0048] 4. Clinical applicability and value: Precise preoperative planning: The high-precision three-dimensional skeletal structure map provided by this invention can help doctors clearly identify anatomical variations before surgery and formulate safer surgical paths.

[0049] Intraoperative "No-fly Zone" Warning: The extremely high edge precision enables this technology to be used to construct a reliable intraoperative navigation "No-fly Zone," which is expected to significantly reduce the probability of accidental injury to key risk structures such as the optic nerve and internal carotid artery during surgery, thereby reducing the occurrence of serious complications.

[0050] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A medical image segmentation system for paranasal sinus bone structure based on Mamba-UNet, characterized in that, It includes a data acquisition module and an image segmentation module; the data acquisition module is used to acquire medical images of the paranasal sinus bone structure and send the medical images of the paranasal sinus bone structure to the image segmentation module; the image segmentation module is used to receive the medical images of the paranasal sinus bone structure, input the medical images of the paranasal sinus bone structure into the hybrid neural network model H-IHF-UNet, and obtain segmented images; the hybrid neural network model includes an encoder and a decoder, the encoder includes multiple hierarchical fusion blocks, and the decoder includes multiple CNN fusion modules.

2. The medical image segmentation system for paranasal sinus bone structure based on Mamba-UNet according to claim 1, characterized in that, Medical images of the paranasal sinus bone structure are input into the hybrid neural network model H-IHF-UNet to obtain segmented images. This includes: inputting the medical images of the paranasal sinus bone structure into the encoder of the H-IHF-UNet hybrid neural network model to obtain encoded images; specifically, inputting the medical images of the paranasal sinus bone structure into a layered fusion block to obtain a first layered image; downsampling the first layered image to obtain a downsampled first layered image; inputting the downsampled first layered image into the layered fusion block to obtain a second layered image; downsampling the second layered image to obtain a downsampled second layered image; inputting the downsampled second layered image into the layered fusion block to obtain a third layered image; downsampling the third layered image to obtain a downsampled third layered image; and inputting the downsampled third layered image into the layered fusion block. The fourth layer image is obtained, and it is downsampled to obtain an encoded image. The encoded image is then passed through the CNN fusion module of the decoder to obtain a segmented image. Specifically, the encoded image is upsampled to obtain the input image of the decoder. The input image of the decoder is concatenated with the third layer image, and the concatenated image is convolved to obtain a first fused image. The first fused image is then upsampled to obtain an upsampled first fused image. The upsampled first fused image is then concatenated with the second layer image, and the concatenated image is convolved to obtain a second fused image. The second fused image is then upsampled to obtain an upsampled second fused image. The upsampled second fused image is then concatenated with the first layer image, and the concatenated image is convolved to obtain a third fused image. The third fused image is then passed through a segmentation layer to obtain a segmented image.

3. A medical image segmentation system for paranasal sinus bone structure based on Mamba-UNet according to claim 2, characterized in that, Medical images of the paranasal sinus bone structure are input into a layered fusion block to obtain the first layer image, including: medical images of the paranasal sinus bone structure. After the average pooling layer, the summary image is obtained. ; The summary graph is processed using a Mamba module that includes a gated bidirectional fusion mechanism. The process is performed to obtain a summary diagram after global modeling. The summary diagram after global modeling processing The global context feature map is obtained by upsampling using bilinear interpolation. , global context feature map Adding the medical images of the paranasal sinus bone structure to obtain a global information feature map. ; global information feature map The H-Mamba module is used to extract local details in a refined manner, resulting in the first layer image.

4. The medical image segmentation system for paranasal sinus bone structure based on Mamba-UNet according to claim 3, characterized in that, The summary graph is processed using a Mamba module that includes a gated bidirectional fusion mechanism. The process is performed to obtain a summary diagram after global modeling. This includes: abstract diagrams Forward features were obtained by performing forward Mamba scans and backward Mamba scans respectively. and backward features ; forward features and backward features The feature maps are concatenated along the channel dimension, and then passed through a fully connected layer and a sigmoid activation function to obtain a gated weight map. ; through gating weight graph Forward features and backward features Perform element-wise weighted summation to obtain the summary graph after global modeling. 。 5. A medical image segmentation system for paranasal sinus bone structure based on Mamba-UNet according to claim 4, characterized in that, Through the gating weight graph Forward features and backward features Perform element-wise weighted summation to obtain the summary graph after global modeling. Specifically, this is achieved through the following formula: 。 6. A medical image segmentation system for paranasal sinus bone structure based on Mamba-UNet according to claim 3, characterized in that, Global information feature map The input to the H-Mamba module is used for refined extraction of local details to obtain the first layer image, including: extracting global information feature maps. After passing through high-frequency, mid-frequency, and low-frequency branches respectively, high-frequency, mid-frequency, and low-frequency feature maps are obtained. The mid-frequency and low-frequency feature maps are upsampled, and then concatenated along the channel dimension to obtain a first concatenated feature map. This first concatenated feature map is then processed... After convolutional layers and global information feature maps Add them together to obtain the first layer image.

7. A medical image segmentation system for paranasal sinus bone structure based on Mamba-UNet according to claim 6, characterized in that, Global information feature map After passing through high-frequency, mid-frequency, and low-frequency branches respectively, high-frequency feature maps, mid-frequency feature maps, and low-frequency feature maps are obtained, including: global information feature maps. After high-frequency branching, a high-frequency feature map is obtained; specifically, this is achieved through the global information feature map. Divide the size by the first size to get the first number of blocks, and then use the first number of blocks to create a feature map of the global information. The data is divided into multiple high-frequency sub-feature maps. These high-frequency sub-feature maps are then processed through a Mamba module containing a gated bidirectional fusion mechanism to obtain the high-frequency feature map. The global information feature map is then processed. After the intermediate frequency branch, the intermediate frequency feature map is obtained; specifically, the global information feature map... Downsampling by a factor of 4 yields a downsampled feature map. Dividing this downsampled feature map by the second size gives the second number of blocks. The downsampled feature map is then divided according to this second number of blocks, resulting in multiple intermediate frequency (IF) sub-feature maps. These IF sub-feature maps are then processed through a Mamba module containing a gated bidirectional fusion mechanism to obtain the final IF feature map. The global information feature map is then... After the low-frequency branch, a low-frequency feature map is obtained; specifically, this is the global information feature map. Downsampling by 16 times yields a downsampled feature map. The downsampled feature map is then divided by the third size to obtain the third number of blocks. The downsampled feature map is then divided according to the third number of blocks to obtain multiple low-frequency sub-feature maps. These multiple low-frequency sub-feature maps are then processed through a Mamba module containing a gated bidirectional fusion mechanism to obtain a low-frequency feature map.

Images