Thalamus subregion segmentation method and system based on 7TMR-PF nucleus specificity
By optimizing a hybrid network architecture and anatomically weighted loss function, and combining the results of micro-expression recognition, the problems of accuracy and computational complexity in thalamic PF nucleus segmentation were solved, achieving efficient and accurate PDD imaging diagnostic support.
Patent Information
- Application Number
- CN202511233807.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-01
- Publication Date
- 2025-12-16
AI Technical Summary
Existing technologies for segmenting the PF nucleus in the thalamus suffer from problems such as difficulty in segmenting small targets, blurred boundaries, lack of pathological correlation, and high computational complexity, resulting in insufficient segmentation accuracy and difficulty in real-time operation on clinical devices.
A thalamic subregion segmentation method based on 7TMR-PF nuclei specificity is adopted, which combines a hybrid network of CNN, visual MLP and axial shift self-attention mechanism. CycleGAN is used to generate 3T-class 7T images, an anatomically weighted loss function is introduced, and the segmentation model is optimized by combining micro-expression recognition results. Accurate segmentation is achieved through a three-level feature extraction architecture.
It improves the segmentation accuracy and computational efficiency of PF nuclei, enhances the correlation between segmentation results and pathological features, is suitable for real-time clinical diagnosis, and provides quantitative indicators to assist in the early diagnosis of PDD.
Smart Images
Figure CN121147237A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image analysis and diagnosis of neurodegenerative diseases, specifically to a method for thalamic subregion segmentation based on 7T magnetic resonance imaging (MRI) and pathological features of the parafascicular nucleus (PF) of the thalamus, and a clinical auxiliary diagnostic system for the early imaging diagnosis of Parkinson's disease with depressive symptoms (PDD). Background Technology
[0002] Precise segmentation of thalamic subregions (especially the PF nucleus) is crucial for the diagnosis of PDD, but it faces the following challenges:
[0003] The challenge of segmenting small targets: The volume of the PF kernel is only about 2-3 mm³. Traditional models such as U-Net suffer from blurred boundaries due to local inductive bias, resulting in a segmentation error rate as high as 19.3%.
[0004] Pathological association missing: Current technology does not combine the pathological mechanism of PF nuclei in the pathogenesis of PDD (Feng Guoping team, 2022), and cannot capture the imaging features of nucleus degeneration.
[0005] High computational complexity: Traditional Transformer models are difficult to run in real time on clinical devices due to their excessive global computational load.
[0006] While 7TMRI can improve the resolution of PF nuclei, existing segmentation methods (such as 3DU-Net) are not optimized for the magnetic susceptibility characteristics of PF nuclei and do not incorporate neuropathological constraints, resulting in segmentation results that lack clinical diagnostic value. Summary of the Invention
[0007] In view of this, in order to solve the problems of insufficient segmentation accuracy and blurred boundaries of small targets in 7TMR images of the thalamus PF nucleus, the present invention provides a thalamic subregion segmentation method and system based on 7TMR-PF nucleus specificity, so as to achieve accurate extraction of PDD-specific imaging markers.
[0008] To achieve the above objectives, one of the present invention provides the following technical solution:
[0009] The thalamic subregion segmentation method based on 7TMR-PF nuclei specificity includes the following steps:
[0010] S1. Construct a hybrid network including CNN, visual MLP and axial shift self-attention mechanism to segment thalamic subregions of 7TMR images;
[0011] S2. Using CycleGAN to synthesize 7T class images of 3TMRI to enhance the magnetic susceptibility characteristics of PF nuclei;
[0012] S3. An anatomically weighted loss function is introduced to assign a pathological weight coefficient of 1.8 to the PF nucleus.
[0013] The beneficial effects of this basic scheme are as follows:
[0014] S1. Local features are extracted through CNN, spatial relationships are modeled by visual MLP, and global dependencies are captured by axial shift self-attention mechanism, which is suitable for the high resolution characteristics of 7TMRI.
[0015] S2. Convert 3T MRI to 7T images to enhance the magnetic susceptibility contrast of PF nuclei (such as iron deposition signals) and solve the problem of insufficient resolution despite the widespread availability of 3T equipment;
[0016] S3. Assigning a 1.8-fold pathological weight to the PF nucleus forces the model to focus on this region, improving segmentation accuracy. This approach is the first to combine cross-modal generative adversarial networks with anatomically weighted loss to achieve specific segmentation of the PF nucleus.
[0017] Furthermore, the encoding layer of the hybrid network uses the ACMix module, which fuses local convolutional features with global self-attention dependencies. The calculation formula is as follows:
[0018]
[0019] Where Q, K, and V represent the feature vectors of the query, key, and value, respectively; It is the dimension of the key vector; Calculate the similarity score; use the softmax function to normalize the score.
[0020] Beneficial effects: While maintaining computational efficiency, it enhances the richness of feature representation, making it particularly suitable for segmentation of multi-scale structures in medical images.
[0021] Furthermore, the axial shift self-attention mechanism reduces computational complexity through the following steps:
[0022] S1. Divide the feature map into blocks along the axis and calculate self-attention within each block;
[0023] S2. Increase the attention weight of the PF nucleus region by 1.5 times.
[0024] Beneficial effects: Improves the segmentation accuracy of PF nuclei without significantly increasing computational load, making it suitable for real-time clinical applications.
[0025] Furthermore, the anatomical weighted loss function is:
[0026]
[0027] in, : Pathological weighting coefficient of PF nuclei; : Real tags; : Model prediction probability.
[0028] Beneficial effects: Enhances the model's ability to identify lesion areas by targeting the pathological changes of PF nuclei in Parkinson's disease (PDD), such as iron deposition and atrophy.
[0029] Furthermore, the segmentation model was optimized by combining the micro-expression recognition results of PDD patients, including: extracting the co-occurrence frequency of AU4 and AU15 in micro-expressions as a functional indicator, and adjusting the attention weight of the PF kernel using the following formula:
[0030]
[0031] in, : The final weight value of the PF kernel; Basic pathology weight; : Activation intensity of AU4-15 FacialActionUnits.
[0032] Beneficial effects: For the first time, facial action units are combined with thalamic nuclei segmentation to achieve joint functional-structural modeling.
[0033] The second aspect of this invention is achieved through the following solution:
[0034] A thalamic subregion segmentation system based on 7TMR-PF nuclei specificity includes:
[0035] The image preprocessing module is used for denoising, resampling, and 7T-class image synthesis of 7T / 3TMR images;
[0036] The hybrid network segmentation module includes ACMix, ASP-MLP, and axially shifted self-attention units;
[0037] The pathology correlation optimization module integrates anatomical weighted loss and micro-expression cross-modal constraints.
[0038] The clinical assessment module outputs the PF nucleus decay index and PDD risk prediction.
[0039] Furthermore, the hybrid network segmentation module adopts a three-level feature extraction architecture:
[0040] Shallow CNNs extract edge details of PF kernels;
[0041] Mid-level visual MLP modeling of the spatial relationship between nuclei and other regions of the thalamus;
[0042] Deep axial shift self-attention enhances the global features of the nucleus.
[0043] Beneficial effects: This system enables the intelligent and precise extraction of PDD-specific imaging markers, making the operation more stable, accurate, efficient, and faster. The entire process from data preprocessing to clinical report generation is automated, simulating the hierarchical processing mechanism of the human visual system. It gradually optimizes the segmentation results from local to global, increasing the correlation between PF nucleus segmentation results and HAMD scale scores by 35%.
[0044] Furthermore, the pathological association optimization module dynamically adjusts the weights of the loss function of the segmentation model by constructing a "PF kernel-micro-expression AU" association matrix.
[0045] Beneficial effects: Quantitatively linking imaging indicators with clinical symptoms enhances the pathological interpretability of segmentation results.
[0046] Furthermore, the clinical assessment module generates a diagnostic report that includes the following indicators:
[0047] The deviation of PF nucleus volume from the normal control group;
[0048] Microexpression-AU co-occurrence anomaly index;
[0049] Probability of PDD incidence based on LASSO regression.
[0050] Beneficial effects: Provides clinicians with quantitative indicators to assist in early diagnosis and treatment planning.
[0051] Furthermore, the image preprocessing module utilizes the Syn_SegNet framework and employs a semi-supervised learning strategy of 7T-guided 3T segmentation.
[0052] Beneficial effect: Robust segmentation of cross-field intensity MRI without relying on a large amount of labeled data.
[0053] The beneficial effects of this invention are as follows:
[0054] 1. Improve the segmentation accuracy of thalamic subregions (especially the PF nucleus).
[0055] Advantages of the hybrid network architecture: Combining the local feature capture capability of CNNs, the spatial relationship modeling capability of visual MLPs, and the global dependency learning of axial shift self-attention, it solves the problems of blurred kernel boundaries and difficulty in distinguishing them from surrounding structures in traditional segmentation methods. The high-resolution characteristics of 7TMR combined with the feature fusion capability of the hybrid network improve the Dice coefficient of sub-region boundary segmentation of PF kernels.
[0056] Enhancement of 7T-class features in 3T images: By synthesizing 7T-class images of 3T images using CycleGAN, the magnetic sensitivity features of the PF nuclei, which were originally blurred in 3TMR, are enhanced, making the segmentation accuracy of low field strength (3T) images close to the level of 7T. This solves the clinical pain point of low adoption rate of 7T equipment and expands the scope of application of the technology.
[0057] 2. Reduce computational complexity and improve the efficiency of high-resolution image processing.
[0058] Axial shift self-attention mechanism: By "calculating self-attention by dividing blocks along the axis", redundant calculations are reduced, and the attention weight of the PF nucleus region is strengthened. While ensuring the segmentation quality of key regions, the processing speed of 7TMR high-resolution images is improved to meet the real-time clinical needs.
[0059] 3. Enhance the correlation between segmentation results and pathological features.
[0060] The role of the anatomical weighted loss function: By assigning a pathological weight of 1.8 to the PF nucleus, the model is forced to focus on optimizing the segmentation error of the nucleus during training, thereby reducing the volume measurement error of the PF nucleus and more accurately reflecting its pathological association with neurodegenerative diseases (such as PDD).
[0061] Cross-modal constraints (micro-expression and image fusion): The co-occurrence frequency and activation intensity of AU4 (frowning) and AU15 (downturned corners of the mouth) in the micro-expressions of PDD patients are introduced. The attention weight of PF nuclei is dynamically adjusted through functional indicators, so that the segmentation results not only depend on structural images, but also associate with disease-specific functional features, thereby improving the sensitivity of the segmentation results to the pathological process of PDD.
[0062] 4. Improve the accuracy and comprehensiveness of clinical assessment.
[0063] Effects of the pathological association optimization module: By constructing a "PF nucleus-microexpression AU" association matrix and dynamically adjusting the weight of the loss function, the correlation between the PF nucleus segmentation results and the HAMD (Hamilton Depression Rating Scale) score is improved, directly enhancing the clinical significance of the segmentation results.
[0064] The clinical assessment module provides multi-dimensional outputs, including indicators such as PF nucleus volume deviation, micro-expression-AU co-occurrence abnormality index, and PDD incidence probability (based on LASSO regression). This enables a one-stop assessment from structural imaging to functional characteristics and disease risk, helping doctors quantify the condition, predict progression, and improve diagnostic decision-making efficiency.
[0065] 5. The practicality and scalability of system-level solutions
[0066] End-to-end workflow integration: The integration of image preprocessing (denoising, resampling, 7T-class synthesis), hybrid network segmentation, pathology optimization, and clinical assessment modules simplifies clinical operation procedures, reduces manual intervention, and makes the technology easier to implement in hospitals.
[0067] Advantages of semi-supervised learning: The image preprocessing module adopts the Syn_SegNet framework and uses a semi-supervised strategy of 7T guiding 3T segmentation. It can still guarantee segmentation accuracy when 3T data annotation is insufficient, reducing the dependence on large-scale labeled data and improving the applicability of the technology in data-scarce scenarios.
[0068] In summary, this invention achieves "high precision, strong pathological correlation, and high clinical applicability" in thalamic subregion (especially the PF nucleus) segmentation through structure-function cross-modal fusion, key nucleus optimization, and efficient computing architecture. This provides reliable imaging technology support for the early diagnosis and disease assessment of neurodegenerative diseases such as PDD, thereby solving the technical problems of this application.
[0069] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description
[0070] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:
[0071] Figure 1 This is a flowchart of the thalamic subregion segmentation method based on 7TMR-PF nuclei specificity of the present invention;
[0072] Figure 2 This is a schematic diagram of the thalamic subregion segmentation system based on the 7TMR-PF nucleus specificity of the present invention. Detailed Implementation
[0073] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0074] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0075] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.
[0076] like Figure 1 As shown, this embodiment provides a thalamic subregion segmentation method based on 7TMR-PF nucleus specificity, and the specific steps are as follows:
[0077] S1. Construct a hybrid network to segment thalamic subregions from 7TMR images.
[0078] The hybrid network incorporates CNN, visual MLP, and axial shift self-attention mechanism. The underlying encoder uses the ACMix module, which fuses local convolutional features with global self-attention dependencies. The calculation formula for the ACMix module is as follows:
[0079]
[0080] Where Q, K, and V are the feature vectors of the query, key, and value, respectively, obtained by projection through a 1×1 convolution. is the dimension of the key vector, used to normalize the similarity score. This formula achieves an effective fusion of local convolutional features and global dependencies of self-attention.
[0081] The axial shift self-attention mechanism reduces computational complexity through the following steps:
[0082] First, the feature map is divided into multiple blocks along an axial direction (e.g., the coronal plane), and self-attention is calculated within each block, reducing the workload of global computation. Second, the attention weights of the PF kernel region are enhanced by 1.5 times to highlight the features of this region and improve segmentation accuracy.
[0083] S2. Synthesizing 7T class images from 3T MRI using CycleGAN.
[0084] CycleGAN is an unsupervised image translation model that achieves the translation of 3TMR images into 7T class images through adversarial training with two generators and two discriminators. During training, the generators attempt to generate realistic 7T class images to deceive the discriminators, while the discriminators strive to distinguish between real 7T images and generated 7T class images. Through repeated training, the generated 3T 7T class images acquire magnetic susceptibility features similar to 7TMR images, enhancing the recognizability of PF nuclei.
[0085] S3. Introducing an anatomical weighted loss function
[0086] The formula for the anatomical weighted loss function is:
[0087]
[0088] in, The pathological weighting coefficient for PF nuclei is 1.8. : Real tags; : Model prediction probability.
[0089] This loss function assigns a higher weight to the segmentation error of the PF kernel during model training, prompting the model to focus on optimizing the segmentation results of this region.
[0090] Furthermore, the segmentation model was optimized by incorporating micro-expression recognition results from PDD patients. The co-occurrence frequencies of AU4 (depressor supercilii contraction) and AU15 (corner of the mouth pull-down) in micro-expressions were extracted as functional indicators, and the attention weights of the PF nuclei were adjusted using the following formula:
[0091] in, : The final weight value of the PF kernel; Basic pathology weight; : Activation intensity of AU4-15 FacialActionUnits.
[0092] This formula correlates the functional indicators of micro-expressions with the pathological weights of PF nuclei, enabling the segmentation model to better reflect the patient's pathological state.
[0093] like Figure 2 As shown, this embodiment also provides a thalamic subregion segmentation system, including the following modules:
[0094] Preprocessing operations such as denoising and resampling are performed on 7TMR and 3TMR images to remove noise and artifacts and unify image resolution. Simultaneously, CycleGAN is used to synthesize 3T-class images of the 7T type, and the Syn_SegNet framework is employed to improve the segmentation performance of the 3T images through a semi-supervised learning strategy that guides 3T segmentation with 7T data.
[0095] Hybrid network segmentation module: Includes ACMix, ASP-MLP, and axially shifted self-attention units, employing a three-level feature extraction architecture. The shallow CNN is used to extract edge details of the PF nucleus, the middle-layer visual MLP models the spatial relationship between the nucleus and other regions of the thalamus, and the deep axially shifted self-attention enhances the global features of the nucleus, achieving accurate segmentation of thalamic subregions.
[0096] Pathological Association Optimization Module: Integrates anatomically weighted loss and micro-expression cross-modal constraints. By constructing a "PF nucleus-micro-expression AU" association matrix, the relationship between PF nucleus segmentation results and micro-expression features is analyzed, and the weights of the loss function of the segmentation model are dynamically adjusted to make the segmentation results more consistent with pathological features.
[0097] Clinical assessment module: Based on the segmentation results, the deviation of the PF nucleus volume from the normal control group is calculated, the micro-expression-AU co-occurrence abnormality index is analyzed, and the incidence probability of PDD is predicted based on the LASSO regression model. A diagnostic report containing these indicators is generated to provide a reference for clinical diagnosis.
[0098] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A thalamic subregion segmentation method based on 7TMR-PF nuclei specificity, characterized in that, Includes the following steps: S1. Construct a hybrid network including CNN, visual MLP and axial shift self-attention mechanism to segment thalamic subregions of 7TMR images; S2. Using CycleGAN to synthesize 7T class images of 3TMRI to enhance the magnetic susceptibility characteristics of PF nuclei; S3. An anatomically weighted loss function is introduced to assign a pathological weight coefficient of 1.8 to the PF nucleus.
2. The thalamic subregion segmentation method based on 7TMR-PF nucleus specificity according to claim 1, characterized in that, The hybrid network's encoder layer uses the ACMix module, fusing convolutional local features with self-attention global dependencies. The calculation formula is as follows: Where Q, K, and V represent the feature vectors of the query, key, and value, respectively; It is the dimension of the key vector; Calculate the similarity score; use the softmax function to normalize the score.
3. The thalamic subregion segmentation method based on 7TMR-PF nucleus specificity according to claim 1, characterized in that, The axial shift self-attention mechanism reduces computational complexity through the following steps: S1. Divide the feature map into blocks along the axis and calculate self-attention within each block; S2. Increase the attention weight of the PF nucleus region by 1.5 times.
4. The thalamic subregion segmentation method based on 7TMR-PF nucleus specificity according to claim 1, characterized in that: The anatomical weighted loss function is: in, : Pathological weighting coefficient of PF nuclei; : Real tags; : Model prediction probability.
5. The thalamic subregion segmentation method based on 7TMR-PF nucleus specificity according to claim 1, characterized in that, The segmentation model was optimized based on the micro-expression recognition results of PDD patients, including: extracting the co-occurrence frequency of AU4 and AU15 in micro-expressions as a functional indicator, and adjusting the attention weights of the PF kernels using the following formula: in, : The final weight value of the PF kernel; Basic pathology weight; : Activation intensity of AU4-15 FacialActionUnits.
6. A thalamic subregion segmentation system based on 7TMR-PF nuclei specificity, characterized in that, include: The image preprocessing module is used for denoising, resampling, and 7T-class image synthesis of 7T / 3TMR images; The hybrid network segmentation module includes ACMix, ASP-MLP, and axially shifted self-attention units; The pathology correlation optimization module integrates anatomical weighted loss and micro-expression cross-modal constraints. The clinical assessment module outputs the PF nucleus decay index and PDD risk prediction.
7. The thalamic subregion segmentation system based on 7TMR-PF nucleus specificity according to claim 6, characterized in that, The hybrid network segmentation module adopts a three-level feature extraction architecture: Shallow CNNs extract edge details of PF kernels; Mid-level visual MLP modeling of the spatial relationship between nuclei and other regions of the thalamus; Deep axial shift self-attention enhances the global features of the nucleus.
8. The thalamic subregion segmentation system based on 7TMR-PF nucleus specificity according to claim 6, characterized in that, The pathological association optimization module dynamically adjusts the weights of the loss function of the segmentation model by constructing a "PF kernel-micro-expression AU" association matrix.
9. The thalamic subregion segmentation system based on 7TMR-PF nucleus specificity according to claim 6, characterized in that, The clinical assessment module generates a diagnostic report that includes the following indicators: The deviation of PF nucleus volume from that of the normal control group; Microexpression-AU co-occurrence anomaly index; Probability of PDD incidence based on LASSO regression.
10. The thalamic subregion segmentation system based on 7TMR-PF nucleus specificity according to claim 6, characterized in that, The image preprocessing module utilizes the Syn_SegNet framework and employs a semi-supervised learning strategy of 7T-guided 3T segmentation.