Cross-attention guided explainable deep transformer model for multi-level classification of rare neurological disorders using MRI images.
Authors
Affiliations (7)
Affiliations (7)
- Department of Computer Science and Engineering, Malla Reddy University, Hyderabad, India.
- Department of Computer Science and Engineering, Vignan's Institute of Engineering for Women, Visakhapatnam, 530046, Andhra Pradesh, India.
- AI&DS Department, Chaitanya Bharathi Institute of Technology, Gandipet, 500075, Hyderabad, India.
- Department of Computer Science and Engineering, Aditya University, Surampalem, India.
- Department of Computer Sciences, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University, P.O. Box 84428, Riyadh, 11671, Saudi Arabia.
- Department of Computer Engineering, College of Computing and Informatics, University of Sharjah, Sharjah, United Arab Emirates.
- Computer Science Department, Faculty of Computers and Information, Qena University, Qena, 83523, Egypt. [email protected].
Abstract
Neurological disorders (ND) impact a significant number of the population all over the world, affecting the brain, spinal cord, and peripheral nerves. These disorders are classified as NeuroDegenerative, NeuroBiological, and NeuroDevelopmental disorders, which vary from common to rare disorders. Traditional deep learning methods often fail to generalize efficiently in these settings and are developed for common or single-label disease classification. Also, the insufficient interpretability in these models can decrease the medical professionals' trust and utilization. Therefore, this study introduces a Cross-Attention Guided Explainable Deep Transformer Model for Multi-Level Classification of Rare Neurological Disorders (CAXDT-MLCRND) model. The proposed model involves the integration of convolutional neural network for local feature representation with transformer-based network to capture global contextual dependencies. Besides, a new cross-attention strategy is applied for enabling dynamic interaction among local and global representations, enabling the model to effectively distinguish visually similar rare disease patterns. Moreover, the derived features are fused with complementary representations from convolutional neural networks within a soft voting-based ensemble model, which integrates a Graph Neural Network (GNN), Deep Belief Network (DBN), and Stacked Autoencoder (SAE) to boost the robustness and generalization of the classification process. Additionally, Bayesian optimization (BO) is leveraged to fine-tune the hyperparameters of the DL approach for improved performance. For assuring clinical reliability, explainability is incorporated using Eigen Class Activation Mapping (EigenCAM), offering visual understanding into model prediction and improving clinical interpretability. The experimental results of the CAXDT-MLCRND system are validated utilizing open access MRI dataset, and the results report the enhanced multi-label classification performance on the simultaneous prediction of co-existing neurological conditions. Therefore, the proposed model is found to be a robust and interpretable approach to assist clinical applications for rare neurological disorder detection in resource-limited environments. Moreover, the source code used in this study is publicly available at: https://github.com/researcher010-debug/NeurologicalDisorderClassification.git .