A method for MRI lumbar vertebra and intervertebral disc segmentation based on two-dimensional and three-dimensional collaborative mutual learning

Abstract

The application discloses a two-dimensional and three-dimensional collaborative mutual learning MRI lumbar vertebra and intervertebral disc segmentation method, and belongs to the technical field of medical image intelligent analysis. After anisotropic resampling, intensity normalization and region cropping are performed on the lumbar vertebra MRI data, the data is input into a two-dimensional segmentation branch and a three-dimensional segmentation branch, cross-dimensional information interaction is realized through bidirectional mapping from a slice to a volume and from a volume to a slice, a dynamic gating mutual learning mechanism is constructed based on confidence, consistency and axial continuity to inhibit low-confidence pseudo-supervised noise, and anatomical topology priori is further introduced on the segmentation result to perform instance separation, segment labeling and topology legality correction on the vertebral body and the intervertebral disc. Experiment proves that the method can significantly improve the segmentation accuracy of complex anatomical structures and effectively solve the boundary blur problem in anisotropic data. The application forms a two-dimensional and three-dimensional collaborative mutual learning segmentation idea, and has innovation and interpretability.

Description

Technical Field

[0001] This invention relates to the field of medical image processing and intelligent assisted diagnosis technology, specifically to an MRI lumbar spine and intervertebral disc segmentation method based on two-dimensional and three-dimensional collaborative learning. Background Technology

[0002] Lumbar spondylolisthesis and intervertebral disc herniation are common degenerative spinal diseases in clinical practice. Magnetic resonance imaging (MRI) can effectively visualize the vertebral body, intervertebral disc, spinal canal, and surrounding soft tissue structures, making it an important tool for clinical imaging evaluation, lesion localization, and grade determination. However, current clinical workflows mainly rely on manual image interpretation and structural drawing, which are time-consuming, subjective, and lack coordination with physicians.

[0003] In existing intelligent diagnostic research, two-dimensional models have certain advantages in representing intra-slice boundary details, but they are insufficient in characterizing cross-slice spatial continuity, easily leading to inter-slice jitter and structural breaks. While three-dimensional models possess the advantage of voxel-level spatial consistency, they are prone to problems such as sparse inter-slice information, blurred fine boundaries, and high computational costs under the anisotropic acquisition conditions common in lumbar MRI. Some solutions have attempted to fuse two-dimensional and three-dimensional models, but most remain at the level of static stitching, simple voting, or mean fusion, lacking the screening of pseudo-supervision credibility and control over cross-dimensional misleading propagation, making the segmentation results difficult to meet the needs of practical applications. Summary of the Invention

[0004] (a) The technical problems to be solved.

[0005] The purpose of this invention is to provide a lumbar MRI segmentation scheme that balances boundary accuracy and clinical interpretability, and to solve the problems of insufficient coordination between two-dimensional and three-dimensional information, lack of topological constraints, and difficulty in directly supporting structured hierarchical diagnosis in the existing technology.

[0006] (ii) Technical solution.

[0007] This invention provides a method for MRI lumbar spine and intervertebral disc segmentation based on two-dimensional and three-dimensional collaborative learning, comprising the following steps:

[0008] S1. Perform data preprocessing on the acquired lumbar MRI body data to obtain standardized input data;

[0009] S2. The standardized input data is input into the two-dimensional segmentation branch and the three-dimensional segmentation branch of the collaborative learning network, respectively. The two-dimensional segmentation branch is used to extract high-frequency texture details in the slice and output the two-dimensional prediction result. The three-dimensional segmentation branch is used to extract voxel-level global spatial context information and output the three-dimensional prediction result.

[0010] S3. Through a bidirectional dimensional mapping mechanism, cross-dimensional mapping and alignment of features are performed between the two-dimensional segmentation branch and the three-dimensional segmentation branch;

[0011] S4. Evaluate the two-dimensional prediction results and the three-dimensional prediction results based on the dynamic gating mechanism, calculate the dynamic gating weights, and perform mutual supervision and filtering of features between the two-dimensional segmentation branch and the three-dimensional segmentation branch according to the gating weights;

[0012] S5. Combine the target loss function to jointly optimize and train the two-dimensional segmentation branch and the three-dimensional segmentation branch, and output the final image segmentation result based on the trained collaborative learning network.

[0013] Furthermore, the present invention includes the following steps:

[0014] Step 1: Data Acquisition and Preprocessing.

[0015] The acquired lumbar MRI body data were subjected to anisotropic resampling, intensity normalization, region clipping, and data enhancement to obtain standardized input data.

[0016] Step 2: Two-dimensional feature extraction and prediction.

[0017] Standardized input data is fed into the two-dimensional segmentation branch to obtain a slice-level two-dimensional segmentation probability map.

[0018] Step 3: 3D feature extraction and prediction.

[0019] The standardized input data is input into the 3D segmentation branch to obtain the block-level 3D segmentation probability volume.

[0020] Step 4: Two-way dimension mapping.

[0021] A bidirectional dimensional mapping relationship between two-dimensional and three-dimensional segmentation branches is established by using the slice-to-volume recombination operator S2V and the volume-to-slice mapping operator V2S.

[0022] Step 5: Calculate the dynamic gating weight.

[0023] Dynamic gating weights are calculated based on the confidence, consistency, and axial continuity of the two-dimensional and three-dimensional segmentation results, and cross-branch mutual guidance information is weighted and filtered.

[0024] Step 6: Joint training of models.

[0025] Under the common constraints of basic supervision loss, confidence-aware mutual conduction loss, and axial topological continuity loss, the two-dimensional segmentation branch and the three-dimensional segmentation branch are jointly trained.

[0026] Step 7: Target reasoning and segmentation.

[0027] The trained model is used to infer from the lumbar MRI data of the test spine and output accurate segmentation results of the vertebral body, intervertebral disc and spinal canal related structures.

[0028] (iii) Beneficial effects.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] This invention innovatively constructs a segmentation framework for collaborative learning between two-dimensional and three-dimensional models. It leverages the advantages of two-dimensional models in extracting intra-slice boundary details while integrating the consistency of three-dimensional models in voxel-level spatial features. By introducing dynamic gating weights, it effectively filters cross-dimensional information and controls misleading propagation. Simultaneously, by combining axial topological continuity loss, it significantly improves upon the shortcomings of traditional methods, such as inter-slice jitter, structural breaks, and blurred fine boundaries, greatly enhancing the segmentation accuracy of various lumbar spine structures. This provides reliable technical support for clinical imaging assessment and structured hierarchical diagnosis. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating the overall process of a two-dimensional and three-dimensional collaborative learning method for MRI lumbar spine and intervertebral disc segmentation proposed in this invention.

[0032] Figure 2 The diagram shows the architecture of ResUNet-2D, ECSM, and CGSF.

[0033] Figure 3 This is a diagram of the HCMA-S three-dimensional segmentation branch structure.

[0034] Figure 4 This is a schematic diagram of TSF, HFF, and the boundary refinement head.

[0035] Figure 5 This is a schematic diagram of bidirectional dimension mapping, C3-Gate, and composite loss.

[0036] Figure 6 The images show a comparison of segmentation results of different algorithms for the same lumbar MRI data in this embodiment of the invention; where (a) is the preliminary segmentation result of the two-dimensional segmentation branch, (b) is the preliminary segmentation result of the three-dimensional segmentation branch, and (c) is the final fusion segmentation result after two-dimensional and three-dimensional collaborative learning. Detailed Implementation

[0037] Reference Figure 1 The specific implementation steps of the present invention are as follows:

[0038] Step 1: Data preprocessing.

[0039] The acquired lumbar MRI body data were anisotropically resampled to unify spatial resolution, followed by intensity normalization, region cropping, and data augmentation techniques to expand sample diversity, ultimately yielding standardized input data.

[0040] Step 2: Generate a two-dimensional segmentation probability map.

[0041] Standardized input data is fed into the 2D segmentation branch to obtain a slice-level 2D segmentation probability map. The 2D segmentation branch uses an improved ResUNet-2D network (e.g., Figure 2 (As shown). The network comprises an edge-conditional spectral modulation structure ECSM and a coherent gated jump-connection fusion structure CGSF.

[0042] For the c-th channel of the intermediate feature F, its edge energy is calculated as follows:

[0043] ,

[0044] Among them, D x and D y Let H and W represent the discrete difference operators in the width and height directions, respectively, and let e represent the height and width of the feature map, respectively. c This represents the average edge intensity of the c-th channel.

[0045] After obtaining channel gating based on edge energy, the frequency band modulation mask is calculated using the following formula:

[0046] ,

[0047] Where, q c This represents the channel gating coefficient generated from edge evidence, where B represents the number of frequency bands, and α... c,b Let I represent the learnable gain of the c-th channel in the b-th frequency band, and let I(ω∈Bb) be the indicator function.

[0048] The CGSF structure is used to control the strength of dimensional skip connection features. Its consistency mapping formula is:

[0049] ,

[0050] Where Ŝ represents the normalized encoder-side skip connection feature, Û represents the normalized decoder-side feature, γ is the learnable scaling parameter, and A is the consistency mapping.

[0051] The final fusion output formula based on consistency mapping is:

[0052]

[0053] Where S represents the skip connection feature at the encoding end, U represents the feature at the decoding end, ⊙ represents element-wise multiplication, and [⋅,⋅] represents the channel concatenation operation.

[0054] Step 3: Generation of the 3D probabilistic volume.

[0055] Standardized input data is fed into the 3D segmentation branch to obtain a block-level 3D segmentation probability volume. The 3D segmentation branch uses an improved HCMA-S network (such as...). Figure 3 (As shown).

[0056] During the encoding phase, the network controls the compression ratios in the in-plane and inter-layer directions through dynamic topology downsampling. At skip connections, the three-dimensional contextual information of the axial, coronal, and sagittal planes is aggregated using the three-dimensional spatial orthogonal fusion structure TSF. During the decoding phase, boundary details are restored using the high-resolution feature fusion structure HFF, and the boundary segmentation is further smoothed at the output using a boundary refinement head (e.g., ...). Figure 4 (As shown).

[0057] Step 4: Two-way dimension mapping.

[0058] By using the slice-to-volume recombination operator S2V and the volume-to-slice mapping operator V2S, a bidirectional dimensional mapping is established between the two-dimensional and three-dimensional branches.

[0059] S2V mapping is defined by the following formula:

[0060] ,

[0061] Among them, P 2D Let R represent a two-dimensional probability volume. S2V This represents the volume-to-slice mapping operator. This represents the recombined 3D supervised graph.

[0062] V2S mapping is defined by the following formula:

[0063] ,

[0064] Among them, P 3D R represents a three-dimensional probability volume. V2S This represents the volume-to-slice mapping operator. This represents a sliced ​​two-dimensional supervised graph.

[0065] Step 5: Calculate the dynamic gating weight.

[0066] Based on the confidence level, consistency, and axial continuity of the two-dimensional and three-dimensional segmentation results (combined with... Figure 5 The gating weights are dynamically calculated to weighted filter cross-branch communication information. The relevant metric formula is as follows:

[0067] ,

[0068] Where, p i Let H(⋅) represent the class probability vector at position i, and let H(⋅) represent the entropy function. This represents the two-dimensional probability graph predicted by the two-dimensional branch at position i. J represents the extraction and alignment of a 2D probability map slice from the prediction results of the 3D branch to position i. i Pi represents the soft crossover ratio (SCVR) between the 2D segmentation result and the corresponding 3D slice, where Pi represents the 3D class probability volume obtained from the 3D branch prediction. Let Q represent the finite difference operator along the depth direction. i This represents a continuity measure obtained based on finite difference in the depth direction.

[0069] The formula for calculating dynamic gating weights is:

[0070] ,

[0071] Where σ(⋅) represents the Sigmoid activation function, β1, β2, and β3 are weight coefficients, and τ1, τ2, and τ3 are the corresponding thresholds.

[0072] Step 6: Joint training of models.

[0073] The network parameters are jointly optimized under the common constraints of basic supervision loss, confidence-aware mutual conduction loss, and axial topological continuity loss.

[0074] The basic supervision loss includes Dice loss and cross-entropy loss, and the calculation formula is as follows:

[0075] ,

[0076] Among them, L dice L represents the loss in Dice. ce λ represents the cross-entropy loss. ce represents the weighting coefficients for cross-entropy loss.

[0077] The confidence-sensing cross-conversion loss is constrained by gating weights and KL divergence, and the calculation formula is as follows:

[0078] ,

[0079] Where V represents the set of voxels, w i KL(⋅) represents the dynamic gating weight at position i. ⋅) represents the Kullback-Leibler divergence.

[0080] The axial topological continuity loss is calculated using the following formula:

[0081] ,

[0082] Among them, ∇ z Let denote the first-order difference operator along the depth direction, and |⋅|1 denote the L1 norm.

[0083] The total loss function of the network is weighted and summed according to the following formula:

[0084] ,

[0085] in, and λ1, λ2(t), and λ3 represent the basic supervision losses for the two-dimensional and three-dimensional segmentation branches, respectively, and are the dynamic and static weighting coefficients for adjusting the proportions of each loss.

[0086] Step 7: Target reasoning and segmentation.

[0087] After the training process is completed, the trained model is used to infer the lumbar MRI data of the target spine and output a fine three-dimensional spatial segmentation result including the vertebral body, intervertebral disc and spinal canal, which provides a basis for subsequent structured diagnosis and lesion grading.

[0088] Example

[0089] This embodiment applies the method of the present invention to 200 real clinical lumbar spine MRI images to conduct segmentation tests on the lumbar spine and intervertebral discs. In the segmentation process, the input volume data is first spatially resampled to 4.4 mm × 0.625 mm × 0.625 mm and intensity normalized. Subsequently, in the feature extraction and segmentation prediction stage, the standardized data is simultaneously fed into the two-dimensional and three-dimensional segmentation branches. In the cross-dimensional collaborative segmentation stage, the system dynamically generates gating weights using calculated Shannon entropy, soft crossover ratio, and depth difference to adaptively correct the initial segmentation results.

[0090] To visually demonstrate the advantages of the collaborative segmentation of this invention, combined with Figure 6 The comparison charts are analyzed. For example... Figure 6 As shown in the two-dimensional segmentation results in (a), although a single-slice network can capture some free edges, due to the lack of spatial context information in the depth dimension, the mask exhibits obvious edge jitter and interlayer breakage when crossing slices; for example... Figure 6 As shown in the 3D segmentation results in (b), although the 3D network maintains the topological integrity of the vertebral arrangement on a macroscopic level, the fine boundaries of the narrow intervertebral discs are severely blurred and details are lost in the information-sparse regions caused by anisotropy. Figure 6As shown in the final fusion segmentation result in (c), after adopting the collaborative mutual guidance method of the present invention, the system effectively eliminated the isolated false positive segmentation voxels between layers, which not only preserved the sensitivity of the fine edges of the two-dimensional plane, but also took into account the consistency of the three-dimensional spatial context, successfully repaired the broken mask and eliminated the edge blurring.

[0091] Test results show that the method of this invention achieves significant improvements in all core segmentation evaluation metrics, with an average Dice coefficient (mDice) of 0.8633. Specifically, for vertebrae with large target volumes, the segmentation Dice coefficient reaches 0.8698; for intervertebral discs with narrow shapes, the fine segmentation Dice coefficient also reaches 0.8444. In comparison, the mask generation mDice of the traditional 3DU-Net method is only 0.6905, and the mDice of the baseline 3D nnU-Net V2 method is 0.8337. It can be seen that the segmentation method of this invention directly eliminates edge jitter and mask breakage phenomena of anatomical structures during cross-slice segmentation from a mathematical perspective, effectively isolates the cross-dimensional propagation of low-quality pseudo-supervision signals, and significantly improves the final three-dimensional segmentation accuracy and system robustness of complex lumbar vertebrae and intervertebral discs.

[0092] The above embodiments of the present invention are merely examples for clearly illustrating the present invention and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively describe all possible implementations here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

[0093] The core innovations of this invention are mainly reflected in the following aspects:

[0094] This invention leverages the complementary advantages of 2D networks in boundary detail extraction and 3D networks in spatial consistency. Starting from the 2D slice features and 3D volumetric features of medical MRI images, it establishes mapping relationships between different dimensions using a bidirectional dimensional mapping operator, achieving collaborative learning between 2D planar edges and 3D spatial context. Based on the cross-dimensional mapping results, it dynamically calculates gating weights using Shannon entropy, soft crossover ratio, and depth difference of 2D and 3D prediction masks, and combines this with axial topological continuity loss to achieve accurate filtering of low-quality pseudo-supervision signals and adaptive correction of cross-slice mask breaks. The segmentation method based on 2D and 3D collaborative learning proposed in this invention overcomes the shortcomings of traditional single-dimensional segmentation models, such as inter-layer jitter, structural breaks, and blurred fine boundaries, from a mathematical perspective, and has high practicality in high-precision segmentation tasks involving complex lumbar spine and intervertebral discs.

Claims

1. A method for MRI lumbar spine and intervertebral disc segmentation based on two-dimensional and three-dimensional collaborative learning, characterized in that, Includes the following steps: S1. Perform data preprocessing on the acquired lumbar MRI body data to obtain standardized input data; S2. The standardized input data is input into the two-dimensional segmentation branch and the three-dimensional segmentation branch of the collaborative learning network, respectively. The two-dimensional segmentation branch is used to extract high-frequency texture details in the slice and output the two-dimensional prediction result. The three-dimensional segmentation branch is used to extract voxel-level global spatial context information and output the three-dimensional prediction result. S3. Through a bidirectional dimensional mapping mechanism, cross-dimensional mapping and alignment of features are performed between the two-dimensional segmentation branch and the three-dimensional segmentation branch; S4. Evaluate the two-dimensional prediction results and the three-dimensional prediction results based on the dynamic gating mechanism, calculate the dynamic gating weights, and perform mutual supervision and filtering of features between the two-dimensional segmentation branch and the three-dimensional segmentation branch according to the gating weights; S5. Combine the target loss function to jointly optimize and train the two-dimensional segmentation branch and the three-dimensional segmentation branch, and output the final image segmentation result based on the trained collaborative learning network.

2. The method according to claim 1, characterized in that, The data preprocessing in step S1 includes: resampling the input raw medical image volume data to the target voxel spacing, and cropping it into image blocks of a fixed size as input to the network.

3. The method according to claim 1, characterized in that, The two-dimensional segmentation branch adopts an improved ResUNet-2D network structure, which includes an edge-conditional spectrum modulation module (ECSM) and a coherent gated skip-connection fusion module (CGSF). The ECSM uses a discrete gradient operator to estimate channel edge energy, generates edge-conditional channel gating, and assigns learnable bandwidth gains to the frequency components after discrete Fourier transform based on the edge-conditional channel gating to enhance boundary-related frequency information and suppress high-frequency noise. The CGSF is used to calculate the consistency mapping between hop connection features at the encoding end and features at the decoding end, and adaptively controls the introduction strength of hop connection features based on the consistency mapping.

4. The method according to claim 1, characterized in that, The 3D segmentation branch employs an improved HCMA-S network structure, which includes a dynamic topology downsampling structure, a three-way spatial orthogonal fusion module (TSF), a high-resolution feature fusion module (HFF), and a boundary refinement head. The dynamic topology downsampling structure controls the compression ratios in the in-plane and inter-layer directions, respectively. The TSF utilizes three parallel depthwise separable convolutional branches to aggregate 3D spatial context information along the axial, coronal, and sagittal planes, respectively. The HFF recovers high-resolution boundary details from the decoding stage. The boundary refinement head further smooths and corrects the output segmentation boundaries.

5. The method according to claim 1, characterized in that, The bidirectional dimension mapping mechanism includes a slice-to-volume recombination mechanism (S2V) and a volume-to-slice extraction mechanism (V2S). The slice-to-volume recombination mechanism (S2V) fills the 2D predicted slices output from the 2D segmentation branch into a sparse volume according to the slice index, and performs linear interpolation and smoothing along the depth direction to generate a 3D probability volume aligned with the output space of the 3D segmentation branch. The volume-to-slice extraction mechanism (V2S) extracts 2D probability slices from the 3D predicted probability volume output from the 3D segmentation branch, according to the corresponding slice index, that are spatially aligned with the 2D predicted slices.

6. The method according to claim 1, characterized in that, The dynamic gating mechanism is a C3-Gate dynamic gating mechanism with joint constraints of confidence, consistency, and continuity. The dynamic gating mechanism comprehensively evaluates the prediction confidence, cross-dimensional consistency, and inter-slice continuity of the two-dimensional and three-dimensional segmentation results, calculates the dynamic gating weight w, and performs weighted filtering on the mutual supervision information between the two-dimensional and three-dimensional segmentation branches according to the dynamic gating weight w, so as to reduce the misleading propagation of low-confidence pseudo-supervision signals.

7. The method according to claim 1, characterized in that, The target loss function includes basic segmentation loss, confidence-aware cross-transduction loss, and axial topological continuity loss. The basic segmentation loss is a weighted average of Dice loss and cross-entropy loss, used to constrain the basic segmentation accuracy of the 2D and 3D segmentation branches. The confidence-aware cross-transduction loss, based on the dynamic gating weight w, constrains the probability distribution difference between the 3D probability volume obtained by S2V recombination of the 2D prediction result and the 3D probability volume output by the 3D segmentation branch. The axial topological continuity loss is achieved by calculating the L1 norm of the first-order gradient difference in the depth direction between the recombinated 3D probability volume and the 3D prediction probability volume, used to constrain the continuity of inter-layer segmentation results and reduce cross-slice mask breaks.

8. The method according to any one of claims 1 to 6, characterized in that, The acquired lumbar MRI body data is lumbar MRI image data, and the final image segmentation result includes the segmented vertebral body instance mask and intervertebral disc instance mask.

Images