Multi-Modal Transformer-Enhanced NAS for Stroke Subtype Classification with Uncertainty Quantification and Federated Learning

Abstract

Despite advances in automated stroke classification using deep learning approaches, current methods face significant challenges in model generalizability across diverse clinical settings, uncertainty quantification, and privacy-preserving collaborative learning. This paper introduces TransFed-StrokeNet, a novel framework that extends neural architecture search for stroke classification by incorporating transformer-based multi-modal fusion, uncertainty-aware learning, and federated optimization strategies. Our approach leverages vision transformers with cross-attention mechanisms to efficiently integrate information from multiple MRI sequences while maintaining sensitivity to both local and global imaging features characteristic of different stroke subtypes.

TransFed-StrokeNet introduces a hierarchical architecture search space that explores optimal combinations of convolutional and transformer blocks, guided by an uncertainty-aware loss function that simultaneously optimizes for classification performance, calibrated confidence estimates, and model parameter efficiency. The framework employs a novel federated neural architecture search algorithm that enables collaborative model development across healthcare institutions without centralizing sensitive patient data, addressing critical privacy concerns in clinical AI deployment.

Our extensive evaluation on a multi-center dataset comprising 7,845 patients from 18 international hospitals demonstrates that TransFed-StrokeNet achieves 98.3% accuracy in differentiating between hemorrhagic and ischemic stroke, with significant improvements in both performance and uncertainty calibration over existing methods. Moreover, the federated learning strategy maintains comparable performance to centralized training while preserving patient privacy. The proposed approach demonstrates remarkable robustness to distribution shifts across imaging protocols, scanner types, and patient demographics, addressing a major barrier to clinical translation of AI-based diagnostic systems.

The framework's ability to automatically design optimized architectures with reliable uncertainty estimates while enabling privacy-preserving collaborative learning represents a significant advancement toward clinically viable AI systems for acute stroke care. By combining state-of-the-art deep learning techniques with practical considerations for real-world clinical deployment, TransFed-StrokeNet provides a comprehensive solution to the challenges of bringing AI-based stroke diagnosis into routine clinical practice across diverse healthcare environments.