SCAN: Salience-Guided Cross-Domain Aggregation Network for Multimodal Medical Image Fusion

Existing multimodal medical image fusion methods often rely on convolutional neural networks (CNNs) for local feature extraction, but fail to model global relationships effectively. Transformer-based approaches address this limitation but are computationally expensive. In this work, we propose the Salience-Guided Cross-Domain Aggregation Network (SCAN) for efficient and high-performance multimodal medical image fusion. SCAN combines the strengths of CNNs and transformers by introducing a novel nested pyramid residual attention (NPRA) module in the encoder for better local feature extraction and adaptive attention to important regions. We also design a salience-guided dual attention (SGDA) module in the decoder to enhance fused features and preserve fine details. Extensive experiments on three multimodal brain datasets show that SCAN outperforms state-of-the-art methods in both qualitative observation and objective assessment. Future work will explore the scalability of SCAN to other medical imaging tasks and its potential for real-time applications.