Multi-Stage Transformer Fusion for Efficient Intracranial Hemorrhage Subtype Classification

Intracranial hemorrhage, a life-threatening condition with diverse subtypes, constitutes a significant portion of global strokes. Precise subtype classification is vital for optimal treatment decisions, prompting the need for efficient computer-aided diagnosis systems due to the challenges of manual review. This paper presents a novel multi-stage vision transformer model, incorporating an efficient three-branch cross-attention mechanism, seamlessly integrating multi-scale contextual information from slice to scan levels. Extensive experiments demonstrate the remarkable performance of our proposed model on three public datasets, surpassing previously state-of-the-art methods in ICH subtypes classification with significantly reduced computational costs.