Enhancing Medical Visual Question Answering with a Multimodal Feature Fusion Network Based on Densely Connected Attention Streams

Medical Visual Question Answering (MVQA) is a specialized task developed from Visual Question Answering (VQA) and is specifically tailored to the medical domain. The goal of MVQA is to enable computers to analyze and interpret medical images, allowing them to answer questions related to the images and provide clinical decision support for healthcare professionals. However, existing research faces significant challenges in terms of insufficient multimodal feature representation and suboptimal feature fusion, which severely limits the accuracy of medical VQA models. To address these challenges, this paper proposes the Multimodal Feature Fusion Network Based on Densely Connected Attention Flow (DCAMNet). The model enhances image feature extraction through meta-learning and mitigates the impact of noisy data on feature fusion by using a convolutional denoising autoencoder. Additionally, the SE-ResNet module is employed to optimize feature weight allocation, while a semantic attention module is introduced to reduce the focus on irrelevant information. The densely connected attention flow-based feature fusion module is designed to resolve issues related to gradient flow loss and insufficient interaction between modalities. Experimental results demonstrate that DCAMNet achieves an overall accuracy of 78.5% and 84.3% on the VQA-RAD and SLAKE datasets, respectively, outperforming existing state-of-the-art methods.