Tubular structure segmentation method based on frequency domain Mama and topology enhancement

Abstract

The invention discloses a tubular structure segmentation method based on frequency domain Mama and topology enhancement, and the method comprises the steps: carrying out the preprocessing of training data, extracting an initial skeleton and a distance map based on the label or structural features of the training data, generating an extended skeleton map based on inverse distance map weighting, and carrying out the subsequent topological constraint modeling and the subsequent supervision signal generation. The expression capability on the connectivity of the tiny branches and the structure is enhanced; constructing a segmentation network fusing a frequency domain Mama module and a difference perception gating attention module, and realizing cross-frequency feature modeling and multilayer semantic alignment; and step-by-step optimization strategy training is adopted, basic segmentation training is firstly carried out, then topology guidance supervision is introduced, and the overall structure consistency and branch connectivity are improved. Compared with the prior art, the method has the advantages that the high efficiency of the model is kept, the segmentation integrity and topological fidelity of the slender structure are remarkably improved, and more reliable technical support is provided for clinical diagnosis and preoperative planning. And constructing a segmentation network fusing a frequency domain Mama module and a difference perception attention mechanism.

Description

Technical Field

[0001] The present invention relates to the fields of medical image processing and artificial intelligence technology, and in particular to a tubular structure segmentation method based on frequency domain Mamba and topology enhancement. Background Art

[0002] Accurate segmentation of tubular structures (such as nerves, blood vessels, and bronchi) in medical images is crucial for aided diagnosis, preoperative planning, and treatment navigation. However, due to their complex tree-like topology, multi-scale branching structure, and similar grayscale imaging characteristics to surrounding tissues, traditional methods have limited performance in preserving structural continuity and detail, and are prone to breakage, missegmentation, and topological errors.

[0003] Existing deep learning methods are primarily based on spatial domain modeling. While they possess strong local feature extraction capabilities, they struggle to capture the long-range dependencies and structural regularities required for tubular structures. In particular, in three-dimensional volumetric images, topological continuity is difficult to ensure with simple pixel-by-pixel reconstruction strategies. The lack of structural constraints often leads to significant errors in models such as small branches and endpoint connections, compromising their clinical reliability. In recent years, frequency domain modeling has gained increasing attention. Studies have shown that the frequency dimension possesses inherent advantages in capturing global geometric patterns and suppressing high-frequency noise, potentially addressing the global perception shortcomings of spatial domain methods. Furthermore, the structured state-space model (Mamba) exhibits broad application prospects compared to CNNs and Transformers, with its linear-time modeling and efficient global modeling capabilities. However, existing applications of the Mamba architecture in medical imaging primarily focus on the spatial domain, leaving a gap in frequency domain modeling and failing to effectively exploit the structural information inherent in frequency domain features.

[0004] Therefore, there is an urgent need for an efficient segmentation method that can integrate global perception and topological connectivity to achieve more accurate and coherent structural modeling and segmentation prediction of complex tubular structures. Summary of the Invention

[0005] In view of this, the present invention proposes a three-dimensional tubular structure segmentation method based on the Mamba network structure of frequency domain modeling. By combining the frequency domain expression capability with the efficient global modeling capability of the Mamba structure and introducing a skeleton topology enhancement mechanism, the segmentation accuracy and topological integrity of slender structures such as nerves and blood vessels are improved, providing structurally reliable technical support for the precise identification and clinical intervention of complex anatomical structures.

[0006] The method of the present invention includes the following steps: first, preprocessing the training images and their labels, using the label information to extract the initial skeleton and distance map, and then generating an extended skeleton map based on an inverse distance weighting strategy. This map serves as a topological supervision signal to explicitly enhance the modeling of branching structures and connectivity relationships. Second, constructing a segmentation network that integrates a frequency-domain Mamba module with a difference-aware gated attention mechanism. This network adopts an integrated backbone structure with two output branches, generating a probabilistic segmentation map of three-dimensional tubular structures and a structured extended skeleton map, respectively. In the backbone network, the frequency-domain Mamba module learns global dependencies through cross-frequency modeling based on the amplitude and phase spectra of the Fourier transform input. The difference-aware gated attention module implements multi-layer semantic alignment, enhancing the ability to express structural boundaries and discriminate features.

[0007] A step-by-step optimization strategy is adopted during the training process. In the first stage, conventional segmentation training is performed, and probability maps and labels are used for pixel-level supervision to establish basic representation capabilities. In the second stage, an extended skeleton map is introduced as topological supervision, and the output of the second branch is combined to perform topological consistency constraints, guiding the network to focus on structural coherence and the complete expression of tiny branches, and finally generating a three-dimensional segmentation result with detail preservation and topological fidelity.

[0008] The beneficial effects of the present invention are as follows:

[0009] The technical solution proposed in this paper has the following beneficial effects: it introduces a frequency-domain modeling mechanism into the structural state space network (Mamba) for the first time, which is used to model the cross-scale geometric features and topological coherence of three-dimensional tubular structures. This significantly improves the model's global representation and detail reproduction capabilities when dealing with complex anatomical structures such as nerves and blood vessels. By introducing the frequency-domain Mamba module, it effectively captures global structural patterns and small branch features, overcoming the limitations of traditional convolutional or Transformer methods in terms of local receptive field and computational complexity, and achieving efficient modeling of three-dimensional structures. Furthermore, the combination of a difference-aware gated attention mechanism enhances the alignment and fusion of multi-scale semantic features, improving feature discriminability and contextual consistency. Specifically, to enhance the topological integrity of the segmentation, the present invention designs a topological enhancement loss based on skeleton expansion and distance weighting, guiding the model to focus on the continuity and branch connectivity of slender structures, thereby strengthening the structural reliability of the segmentation results. This method strikes a balance between accuracy, topological consistency, and computational efficiency, possessing good clinical applicability. It can provide reliable support for the identification of key structures in scenarios such as vascular decompression and preoperative tumor planning, and has broad application prospects and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the specific implementation of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific implementation or the description of the prior art. Obviously, the drawings described below are an implementation of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0011] Figure 1 This is an overall flow chart of a three-dimensional tubular structure segmentation method of a Mamba network structure based on frequency domain modeling according to an embodiment of the present invention. DETAILED DESCRIPTION

[0012] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments.

[0013] Figure 1 This is an overall flow chart of a three-dimensional tubular structure segmentation method of a Mamba network structure based on frequency domain modeling according to an embodiment of the present invention. The method includes steps S101, S102, and S103.

[0014] Step S101: Preprocess the training data to generate a topology-guided extended skeleton representation. Preprocess the training images and their corresponding labels. Exemplarily, the label map serves as a first constraint for generating two supervisory signals: a first constraint for pixel-level segmentation supervision and a second constraint for constructing the extended skeleton representation required for topology supervision.

[0015] In one embodiment, generating the second constraint basis includes the following sub-steps: extracting a centerline skeleton structure image of the target structure using a morphological method based on the first constraint basis; performing a morphological dilation operation on the centerline skeleton structure image to expand the spatial coverage of the skeleton line to alleviate the training instability caused by the sparse supervisory signal; calculating the Euclidean distance transform of the first constraint basis to obtain a distance transform image, and further negating it to obtain an inverse distance weighted image to enhance the response near the edge area; performing an element-by-element product of the expanded skeleton structure image and the inverse distance weighted image to obtain an expanded skeleton structure representation based on inverse distance weighting. This representation has stronger edge response and center connectivity expression capabilities, and serves as a supervisory signal for topological structure modeling, i.e., the second constraint basis.

[0016] Step S102: Construct a segmentation network that integrates the frequency domain Mamba module and the difference-aware gated attention module. In this step, a deep neural network based on an encoder-decoder architecture is designed to jointly model the semantic region information and topological connectivity characteristics of the target structure. Exemplarily, the segmentation network includes a trunk feature extraction path and two output branches, which are used to generate a first prediction result (i.e., a probabilistic segmentation map) of the tubular structure and a second prediction result (i.e., an extended skeleton structure representation) of the topological enhancement, respectively, to achieve collaborative modeling of structure and topology.

[0017] In one embodiment, the frequency domain Mamba module includes the following sub-steps: using the residual block structure unit to perform deep representation learning on the local spatial feature representation of the input to extract a more discriminative residual feature representation; applying a fast Fourier transform (FFT) to the residual feature representation to separate and obtain an amplitude representation and a phase representation respectively; sending the amplitude representation and the phase representation into a shared Mamba structure module for parallel modeling, aiming to jointly capture the topological continuity in the low-frequency components and the structural details in the high-frequency components; restoring the output of the Mamba module to the spatial domain feature representation through an inverse Fourier transform (IFFT), and fusing it with the original residual feature representation through a residual connection mechanism to achieve efficient interactive fusion of spatial domain and frequency domain features.

[0018] It should be noted that the Mamba operation in the frequency domain Mamba module in this embodiment is a shared module, which models the amplitude representation and phase representation in parallel and shares the same parameters, thereby improving the network's ability to capture the global consistency and detailed features of the target tubular structure.

[0019] In one embodiment, the difference-aware gated attention module includes the following steps: performing convolution operations on the first encoding feature output by the encoder and the first decoding feature of the current layer of the decoder to extract two sets of feature representations, namely, the second encoding feature and the second decoding feature; calculating the difference between the second encoding feature and the second decoding feature to generate a first difference significance feature; fusing the first difference significance feature, the second encoding feature and the first gating feature of the previous scale to generate a first fused feature; processing the first fused feature using convolution and normalization operations to generate a gated weight representation; weighting the second encoding feature using the gated weight representation to obtain a second gated feature; outputting the second gated feature to guide the fine-grained transmission and fusion of the first decoding feature to improve multi-level semantic consistency.

[0020] It should be noted that the second prediction result will be output in the difference-aware gated attention module of the last layer of the network to guide the supervised learning of the topological structure.

[0021] Step S103: Training the segmentation network using a step-by-step optimization strategy. In this step, to improve the network's ability to model the semantic accuracy and topological coherence of tubular structures, a step-by-step optimization training strategy is employed to train the constructed segmentation network in stages. This training strategy consists of a first stage, which establishes basic segmentation capabilities, and a second stage, which enhances topological structure modeling capabilities.

[0022] In one embodiment, the training strategy includes the following sub-steps: in the first stage, the first prediction result and the first constraint basis are supervised for training, a structural segmentation loss function is constructed, the network is trained to perform initial semantic modeling of the target structure area, and basic segmentation representation capabilities are established, and the loss function includes but is not limited to cross entropy loss, Dice loss or a combination thereof; in the second stage, the network parameters learned in the first stage are kept as the initial state, and the second prediction result and the second constraint basis are further introduced, and supervised training is performed through the topological loss function to optimize the network's connectivity, branch consistency and overall topological morphology modeling capabilities of the target structure; by jointly optimizing the above two-stage training objectives, a three-dimensional tubular structure prediction result with high-precision semantic segmentation capabilities and complete topological structure expression capabilities is finally obtained.

[0023] It should be noted that the second stage of training in this embodiment, while maintaining the basic capabilities of the first stage, guides the network to gradually learn topological constraints, effectively improving the robustness to tiny branches, broken connections and structural consistency.

[0024] Finally, it should be noted that the above-described embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them, and the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field can still modify the technical solutions described in the aforementioned embodiments or easily think of changes within the technical scope disclosed by the present invention, or make equivalent replacements for some of the technical features therein within the technical scope disclosed by the present invention; and these modifications, changes or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be based on the scope of protection of the claims.

Claims

1. A tubular structure segmentation method based on frequency domain Mamba and topology enhancement, characterized in that: The following steps are involved: Preprocessing the training images and their labels, using the labels as a first constraint, extracting an initial centerline skeleton structure representation and a distance transform image based on the labels, and generating an extended skeleton structure representation based on inverse distance graph weighting. The extended skeleton structure representation serves as a second constraint for supervised modeling of the topological structure. Construct a segmentation network that integrates the frequency-domain Mamba module and the difference-aware gated attention module. The segmentation network includes a backbone network and two output branches, which respectively output a probabilistic segmentation map of a tubular structure as a first prediction result and a structured extended skeleton map as a second prediction result. A step-by-step optimization training strategy is employed. In the first phase, pixel-level segmentation supervision is performed using the first prediction result and the first constraint basis to establish basic feature representation capabilities. In the second phase, the second prediction result and the second constraint basis are introduced for topological structure supervision, optimizing overall structural consistency and branch connectivity. The final output is a 3D tubular structure segmentation result with micro-branch integrity and topological coherence.

2. The method according to claim 1, characterized in that The first constraint is based on a segmentation label, and the second constraint is based on an extended skeleton graph generated based on the segmentation label, which is used to model the topological connectivity of the target structure.

3. The method according to claim 1, characterized in that The second constraint basis is generated by the following steps: extracting the centerline skeleton structure representation of the first constraint basis, and performing a morphological dilation operation on the skeleton structure representation to obtain an expanded skeleton structure representation to alleviate the sparsity of the skeleton supervision signal; obtaining a distance transformation image for the first constraint basis, and inverting the distance transformation image to obtain an inverse distance weight image to increase the edge weight; performing element-by-element multiplication of the expanded skeleton structure image and the inverse distance weight image to generate a second constraint basis for topological structure constraint modeling.

4. The method according to claim 1, wherein The segmentation network is an encoder-decoder structure and includes a frequency-domain Mamba module and a difference-aware gated attention module to enhance the structural feature expression capability and cross-level semantic consistency, respectively.

5. The method according to claim 4, characterized in that The frequency domain Mamba module includes: a residual block for learning local spatial features; a residual block for performing fast Fourier transform on the residual features to extract amplitude representation and phase representation respectively; the amplitude representation and phase representation are fed into a shared frequency domain Mamba module to model structural dependencies in the frequency domain; the Mamba output is restored to the spatial domain through an inverse Fourier transform and residually connected with the original residual features to achieve interactive fusion of the space-frequency domain.

6. The method according to claim 5, characterized in that The Mamba module is a shared module that achieves joint optimization of low-frequency topological continuity and high-frequency structural details by parallel modeling of amplitude and phase representations, thereby improving the global consistency expression capability of the tubular structure.

7. The method according to claim 4, characterized in that The difference-aware gated attention module includes: performing a convolution operation on the first encoding feature output by the encoder and the first decoding feature of the current layer of the decoder to extract two sets of feature representations, namely, the second encoding feature and the second decoding feature; calculating the difference between the second encoding feature and the second decoding feature to generate a first difference significance feature; fusing the first difference significance feature, the second encoding feature and the first gated feature of the previous scale to generate a first fused feature; processing the first fused feature using convolution and normalization operations to generate a gated weight representation; and weighting the second encoding feature using the gated weight representation to obtain a second gated feature.

8. The method according to claim 1, characterized in that The training strategy is a two-stage training strategy, including: in the first stage, the first prediction result and the first constraint basis are supervised and trained to obtain basic segmentation capabilities; in the second stage, the second prediction result and the second constraint basis are introduced to perform topological structure supervision to optimize the model's expression ability of target connectivity and detail consistency.

9. The method according to claim 1, characterized in that The segmentation network has two output branches, which output a segmentation probability map and an extended skeleton map respectively, and receives segmentation supervision and topology supervision respectively during the training process to achieve coordinated optimization of structural integrity and connectivity.

Images