Related Experiment Video
Updated: Jun 29, 2026
14:27
Identification of Disease-related Spatial Covariance Patterns using Neuroimaging Data
Published on: June 26, 2013
A Dual-Branch Frequency-Aware Attention Framework for Rare Neurological Disease Classification from Brain MRI.
Madallah Alruwaili1, Mahmood A Mahmood2
1Department of Computer Engineering and Networks, College of Computer and Information Sciences, Jouf University, Sakaka 72441, Aljouf, Saudi Arabia.
Diagnostics (Basel, Switzerland)
|June 12, 2026
Summary
RareNeuroXNet, a novel deep learning framework, enhances rare neurological disease diagnosis from brain MRIs by integrating global, local, and frequency features. This approach improves classification accuracy and model calibration for better diagnostic support.
Area of Science:
- Medical Imaging and Radiology
- Artificial Intelligence in Medicine
- Neurology and Neuroscience
Background:
- Diagnosing rare neurological diseases from brain MRI is challenging due to low prevalence and data limitations.
- Deep learning shows promise for image-level recognition but requires careful validation, especially with incomplete patient data.
Purpose of the Study:
- To introduce RareNeuroXNet, a frequency-aware multi-branch attention framework for rare neurological disease classification from brain MRI.
- To evaluate if combining global, local, and frequency-domain MRI representations enhances classification performance, calibration, and interpretability.
Main Methods:
- Developed RareNeuroXNet with three branches: global (whole-image), local (regional features), and frequency (FFT magnitude).
- Employed CBAM attention for feature refinement, fused features, and used a fully connected head for classification.
- Evaluated on a curated dataset using five-fold cross-validation, ablation studies, calibration metrics, and Grad-CAM visualization.
Main Results:
- RareNeuroXNet achieved high internal benchmark performance: accuracy 0.9924, macro F1 0.9924, macro AUROC 0.9998, macro AUPR 0.9992.
- Demonstrated favorable calibration (ECE 0.0052, NLL 0.0276) and outperformed a local-only DenseNet121 model.
- Ablation analysis indicated the local branch as the primary contributor, with FFT and CBAM offering supportive refinement.
Conclusions:
- RareNeuroXNet shows strong, interpretable performance for image-level rare neurological disease classification with stable validation.
- The model exhibits high discrimination and favorable calibration, suggesting potential for clinical decision support.
- Future work requires patient-level, multi-center external validation and 3D multimodal MRI analysis for robust generalization.
