Medical Visual Question Answering (VQA-Med) is a challenging task that involves answering clinical questions related to medical images. However, most current VQA-Med methods ignore the causal correlation between specific lesion or abnormality features and answers, while also failing to provide accurate explanations for their decisions. Moreover, VQA-Med methods suffer from the common language bias problem in generic VQA. To explore the interpretability and language bias of VQA-Med, this paper proposes a novel CCIS-MVQA model for VQA-Med based on a counterfactual causal-effect intervention strategy. This model consists of the modified ResNet for image feature extraction, a GloVe decoder for question feature extraction, a bilinear attention network for vision and language feature fusion, and an interpretability generator for producing the interpretability and prediction results. The proposed CCIS-MVQA introduces a layer-wise relevance propagation method to automatically generate counterfactual samples for improving interpretability and alleviating language bias. Additionally, CCIS-MVQA applies counterfactual causal reasoning throughout the training phase to enhance interpretability and generalization. Extensive experiments on three benchmark datasets show that the proposed CCIS-MVQA model outperforms the state-of-the-art methods. Enough visualization results are produced to analyze the interpretability and debasing performance of CCIS-MVQA.

The blurring boundaries and large variations of digestive polyps with complicated anatomical environments severely degrade the performance of automatic polyp segmentation. Though current deep learning-based polyp segmentation methods achieve competitive results, most of these methods often face significant challenges when dealing with out-of-distribution data and real-time performance in clinical applications. In this study, we propose a new lightweight spatial detail fused context reconstruction network (SDCRNet) for real-time polyp segmentation aiming to obtain satisfactory segmenting accuracy while avoiding large-scale computation. Our SDCRNet builds a novel lightweight ConvPFormer block as MetaFormer encoder baseline and then constructs a new multi-resolution spatial detail fused context reconstruction (SDCR) to effectively reconstruct the lost textures and spatial detail features through layer-by-layer interaction of the shallow spatial details with the deep global context, which avoids maintaining high-resolution network propagation while capturing accurate local information and boundary features. Extensive experiments on four benchmark datasets show our SDCRNet obtains satisfactory segmenting results and real-time performance compared with the state-of-the-art methods.