MADS-VQA: MacroMamba and Dual-Sequence Feature Fusion for Medical Visual Question Answering

Abstract

Medical Visual Question Answering (Med-VQA) aims to generate answers from medical images and natural language questions, supporting clinical decision-making and intelligent healthcare. However, existing methods often rely on homogeneous semantic modeling, limiting their ability to capture global anatomical structures and local pathological details, thereby hindering fine-grained cross-modal alignment. To address these limitations, we propose MADS-VQA, a Med-VQA framework integrating a MacroMamba-based visual encoder and a Dual-Sequence Feature Fusion (DSFF) module. MacroMamba captures long-range dependencies and multi-scale contextual information, while DSFF enables fine-grained cross-modal fusion through spatial alignment, channel re-weighting, and sequence reorganization. Experiments demonstrate that MADS-VQA achieves competitive or superior performance across multiple medical VQA benchmarks, improving accuracy by 4.01 percentage points over the strong baseline PitVQA++ on EndoVis-18-VQA while delivering competitive results on Med-VQA (C1–C3). These findings highlight its enhanced multi-scale representation, cross-modal alignment, and robustness across diverse medical imaging scenarios.