Papers

BESR 提出双侧证据选择与推理框架,提升乳腺钼靶 VLM 诊断性能

Bilateral Evidence Selection and Reasoning for Mammography Vision--Language Models

Why we picked it

医学影像 VLM 压缩 token 时最容易丢掉诊断线索,这篇论文把对侧乳房当参照物,用选择器加奖励机制让模型真正会用对比证据,在 CBIS-DDSM 和 VinDr-Mammo 上都验证了效果,还降了 prefill 延迟。做医疗多模态或 token 压缩的值得细读。

论文提出 BESR(Bilateral Evidence Selection and Reasoning),针对医学视觉-语言模型在高分辨率乳腺钼靶影像中识别微小病灶的难题。方法包含两个核心组件:bilateral-guided token selector 在共享 token 预算下同时优先保留病灶外观、局部上下文和对侧乳房的 bilateral 上下文,以及一个 bilateral reward,奖励模型利用保留的对侧对比证据做出诊断评估。

Body

实验在基于 CBIS-DDSM 构建的双侧数据集和 VinDr-Mammo 上进行,结果显示在受限视觉 token 预算下,证据保留率和诊断性能均有提升。推理基准测试还显示 prefill 延迟降低,说明该方法在压缩视觉 token 的同时兼顾了效率和准确性。

论文的出发点是现有视觉 token 压缩只保留显著区域会丢弃诊断解读所需上下文,而对侧乳房可作为患者特异性的解剖参照。BESR 把证据保留与对比评估绑定起来,解决了保留证据不等于有效使用证据的问题。

Original · arXiv cs.AI

Bilateral Evidence Selection and Reasoning for Mammography Vision--Language Models

Medical vision--language models (VLMs) often need to identify small, subtle lesions within high-resolution images containing extensive normal anatomy and redundant visual information. Visual-token compression improves efficiency, but preserving salient regions alone can discard the context needed for diagnostic interpretation. This context can extend across images: in bilateral mammography, the opposite breast provides a patient-specific anatomical reference. Yet retaining such evidence does not ensure that the model uses it effectively. We propose Bilateral Evidence Selection and Reasoning (BESR), which connects evidence preservation with comparative assessment through a bilateral-guided token selector and a bilateral reward. Under a shared token budget, the selector jointly prioritizes lesion appearance and its local and bilateral context, while the reward encourages assessments that benefit from the preserved comparison. Experiments on a bilateral dataset constructed from CBIS-DDSM and on VinDr-Mammo demonstrate improved evidence retention and diagnostic performance under constrained visual-token budgets. Inference benchmarking further shows reduced prefill latency.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk. Key entities and technical topics under continuous tracking include VLM, medical imaging, CBIS-DDSM, VinDr-Mammo.

Source attribution: curated from arXiv cs.AI (@Zhengbo Zhou, Dooman Arefan, Shandong Wu).