← Back to Computer Vision
cs.CV

How to help AI doctors spot tiny, hidden lesions without retraining

Qiwei Zeng, Hao Wang, Jinghao Lin, Shuchang Ye, Yuezhe Yang, Yige Peng, Haoyuan Che, Jinman Kim, Lei Bi

June 4, 2026

Medical vision-language models struggle with tiny, low-contrast lesions because weak visual cues get buried when combining local image patches into global representations. EasyLens adds a training-free layer that identifies lesion-relevant patches using pathology prototypes, then amplifies their signal in the final embedding—like turning up the volume on faint acoustic signals. Works with any frozen medical VLM across multiple imaging datasets without retraining.
Published as EasyLens: A Training-Free Plug-and-Play Subtle-Lesion Representation Amplifier for Medical Vision-Language Models arXiv:2606.06379
Read the original paper →