A HybridConformerNet network to identify structural biomarkers in the dystonia
Poster No:
89
Submission Type:
Abstract Submission
Authors:
Xu Jinping1, Qinxiu Cheng2, Lin Wang3, Gang liu4
Institutions:
1Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, shenzhen, Guangdong, 2Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, shenzhem, Guangdong, 3Aviation General Hospital, China Medical University, shenzhen, Guangdong, 4Department of Neurology, The First Affiliated Hospital, Sun Yat-sen University, Guangdong Provincial, shenzhen, Guangdong
First Author:
Xu Jinping
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
shenzhen, Guangdong
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
shenzhen, Guangdong
Co-Author(s):
Qinxiu Cheng
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
shenzhem, Guangdong
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences
shenzhem, Guangdong
Gang liu
Department of Neurology, The First Affiliated Hospital, Sun Yat-sen University, Guangdong Provincial
shenzhen, Guangdong
Department of Neurology, The First Affiliated Hospital, Sun Yat-sen University, Guangdong Provincial
shenzhen, Guangdong
Introduction:
Isolated dystonia is a neurological disorder marked by abnormal muscle contractions, leading to significant functional impairment.[1] Diagnosis remains challenging due to its reliance on subjective clinical evaluations and the absence of reliable biomarkers.[2] Recent studies suggest structural brain alterations, particularly in the basal ganglia and sensorimotor networks,[3] could serve as potential biomarkers. In this study, we proposed the use of a novel deep learning architecture, HybridConformerNet, to identify structural biomarkers associated with isolated dystonia using structural magnetic resonance imaging (sMRI). The model integrated Convolutional Neural Networks[4][5] for local feature extraction and Vision Transformers[6] for global contextual analysis. The approach achieved an accuracy of 0.97 on the training dataset and 0.96 on the test dataset in two-way classification. To gain insights into the model's predictions, we employed techniques for interpreting deep learning models, identifying four key clusters of structural biomarkers:(1). right inferior temporal gyrus/right cerebellar. (2).bilateral middle cingulate cortex /supplementary motor area. (3).left precentral gyrus. (4).left fusiform gyrus. These findings offer a robust solution for the early diagnosis of this complex disorder and contribute to advancing our understanding of its underlying structural brain changes.
Methods:
We collected sMRI data from three centers. First Affiliated Hospital of Sun Yat-sen University (SYU): 74 patients with BSP, 31 with BOM, 42 with CD, and 136 HCs were included. Aviation General Hospital of China Medical University (AGH)：31 patients with BSP, 92 with BOM, and 36 with CCD were included in the second center, Additionally, we obtained T1-weighted MRI images of 405 HCs from the IXI database.
Preprocessing: T1-weighted images were segmented into gray matter, white matter, and CSF, then transformed into MNI space and cropped.
Data Alignment: VQ-VAE model aligned gray matter images from different centers to a common space, ensuring robustness.
Model Framework
A hybrid framework combining 3D CNNs (local feature extraction) and ViTs (modeling long-range dependencies) was developed. The model used a linear attention mechanism for efficiency.
3D CNN Branch: Extracts local features using convolution layers and max-pooling, followed by fully connected layers for feature vector generation. Dropout and batch normalization prevent overfitting.
Vision Transformer (ViT) Branch: Uses a patch-embedding module to process images into non-overlapping patches, with DenseNet for extracting 512 local features. Applies spatial reduction and feature transformation for improved efficiency. The workflow is illustrated in Fig. 1.
Preprocessing: T1-weighted images were segmented into gray matter, white matter, and CSF, then transformed into MNI space and cropped.
Data Alignment: VQ-VAE model aligned gray matter images from different centers to a common space, ensuring robustness.
Model Framework
A hybrid framework combining 3D CNNs (local feature extraction) and ViTs (modeling long-range dependencies) was developed. The model used a linear attention mechanism for efficiency.
3D CNN Branch: Extracts local features using convolution layers and max-pooling, followed by fully connected layers for feature vector generation. Dropout and batch normalization prevent overfitting.
Vision Transformer (ViT) Branch: Uses a patch-embedding module to process images into non-overlapping patches, with DenseNet for extracting 512 local features. Applies spatial reduction and feature transformation for improved efficiency. The workflow is illustrated in Fig. 1.
·Flowchart of the Proposed Framework. (A) Data Preprocessing Steps for sMRI Data: 1) Head motion correction, 2) Skull stripping, 3) Segmentation into gray matter, white matter, and cerebrospinal fluid,
Results:
The results revealed that our deep learning model accurately classified dystonia in multicenter sMRI scans (Fig. 2). The proposed method outperformed other 3D methods in almost all metrics. In SYU, the ACC values were 0.960/0.863/0.903, AUC values were 0.978/0.960/0.972, recall values were 0.958/0.855/0.896, precision values were 0.980/0.893/0.924, and F1-scores were 0.938/ 0.872/0.907 for the three tasks, respectively. In AGH, the recall values were 0.977, 0.982, and 0.977 for the three tasks, respectively, while the precision was 0.950. Our approach demonstrated a superior balance between high performance and stability, thus enhancing its utility in clinical settings where reliability is crucial. All the results of the proposed method surpassed those of the other models, which demonstrate the optimal generalizability of the proposed model.
·The accuracy of three centers
Conclusions:
Four clusters were identified in the model interpretation, including the right inferior temporal gyrus (ITG.R)/right cerebellum (CERE.R), bilateral middle cingulate cortex (MCG)/supplementary motor area (SMA), the left precentral gyrus (PreCG.L), and the left fusiform gyrus (FG.L)
Disorders of the Nervous System:
Neurodegenerative/ Late Life (eg. Parkinson’s, Alzheimer’s) 1
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
Keywords:
MRI
Neurological
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
2.Kaur M, Sharma U, Solanki RK. Anesthetic nuances in Segawa's syndrome: A case report and review of the literature. Saudi J Anaesth. Oct-Dec 2020;14(4):524-527.
3.Gao, Z., et al., Activation likelihood estimation identifies brain regions activated during puncturing at Hegu in healthy volunteers: A meta-analysis. Front Neurosci, 2022. 16: p. 1084362.
4.He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition, 2016.
5.Valeriani D, Simonyan K. A microstructural neural network biomarker for dystonia diagnosis identified by a DystoniaNet deep learning platform. Proc Natl Acad Sci U S A. Oct 20 2020;117(42):26398-26405.
6.Vaswani A. Attention is all you need. Advances in Neural Information Processing Systems. 2017.