Less Is More? Selective Visual Attention to High-Importance Regions for Multimodal Radiology Summarization

2026-03-31

AI Summary

This study addresses the susceptibility of existing medical imaging report summarization models to visual noise when generating “impression” sections from “findings,” which often prevents them from outperforming purely text-based approaches. Challenging the assumption that more visual input is inherently better, this work proposes a multi-stage selective visual fusion framework. It leverages MedSAM2 for lung segmentation to localize pathological regions and integrates bidirectional cross-attention for multi-view fusion, Shapley-value-guided adaptive image patch clustering, and a hierarchical visual tokenization strategy within a Vision Transformer. By selectively utilizing only high-importance visual regions, the method achieves state-of-the-art performance on MIMIC-CXR (BLEU-4: 29.25%, ROUGE-L: 69.83%), significantly enhancing factual consistency, clinical relevance, and human expert ratings.

Abstract

Automated radiology report summarization aims to distill verbose findings into concise clinical impressions, but existing multimodal models often struggle with visual noise and fail to meaningfully improve over strong text-only baselines in the FINDINGS $\to$ IMPRESSION transformation. We challenge two prevailing assumptions: (1) that more visual input is always better, and (2) that multimodal models add limited value when findings already contain rich image-derived detail. Through controlled ablations on MIMIC-CXR benchmark, we show that selectively focusing on pathology-relevant visual patches rather than full images yields substantially better performance. We introduce ViTAS, Visual-Text Attention Summarizer, a multi-stage pipeline that combines ensemble-guided MedSAM2 lung segmentation, bidirectional cross-attention for multi-view fusion, Shapley-guided adaptive patch clustering, and hierarchical visual tokenization feeding a ViT. ViTAS achieves SOTA results with 29.25% BLEU-4 and 69.83% ROUGE-L, improved factual alignment in qualitative analysis, and the highest expert-rated human evaluation scores. Our findings demonstrate that less but more relevant visual input is not only sufficient but superior for multimodal radiology summarization.

Problem

• radiology summarization
• multimodal learning
• visual attention
• medical image analysis
• report generation

Innovation

• selective visual attention
• multimodal summarization
• Shapley-guided patch clustering
• ViTAS
• radiology report generation