EasyLens: A Training-Free Plug-and-Play Subtle-Lesion Representation Amplifier for Medical Vision-Language Models

NeurIPS, 2026

Hao Wang^1,2,3*, Qiwei Zeng^1,3*, Jinghao Lin^4*, Shuchang Ye², Yuezhe Yang^1,2, Yige Peng¹, Haoyuan Che^3†, Jinman Kim^2†, Lei Bi^1†

^*Equal Contribution ^†Corresponding author

¹Shanghai Jiao Tong University, ²The University of Sydney, ³Jilin University, ⁴Northeastern University

Abstract

Medical vision-language models (VLMs) have shown increasing potential for clinical image interpretation, including lesion detection and report generation. However, their practical utility remains limited by insufficient sensitivity to subtle lesions, whose visual evidence is often sparse, low-contrast, and embedded within complex anatomical context. As local visual tokens are aggregated, these weak lesion cues can become underrepresented in global image representations, making them difficult for medical VLMs to recognize. Existing efforts to improve lesion sensitivity mainly rely on medical-domain vision-encoder pre-training, clinical-term-guided alignment, or trainable pathological representation enhancement. Although effective, these approaches usually require additional training or model-specific adaptation and may overfit to particular disease morphologies, limiting their applicability to frozen medical VLMs. To address these limitations, we propose EasyLens, a training-free plug-and-play subtle-lesion representation amplifier for medical VLMs. EasyLens first constructs EasyBank, a pathology-anatomy prototype space that provides lesion-related prototypes and anatomy-aware normal references for comparing suspicious patches against both pathological and normal anatomical patterns. To avoid blindly amplifying normal tissues, EasyTag selects lesion-relevant patches through counterfactual prototype reasoning. To counteract the dilution of subtle lesion cues in global image representations, EasyAmplifier strengthens the selected lesion-relevant patch representations through morphology-guided residual enhancement, thereby increasing their contribution to the global image embedding. Experiments on multiple medical image datasets and frozen medical VLM backbones show that EasyLens consistently improves subtle-lesion detection and outperforms existing encoder-enhancement baselines without model fine-tuning. Code is available at: https://github.com/wanghao1118/EasyLens

Highlights

Build pathology-anatomy prototypes for comparing suspicious image patches.
Select lesion-relevant patches and amplify subtle morphological cues.
Improve subtle-lesion recognition without fine-tuning frozen medical VLMs.

Results

Quantitative Results

Model	ReX			LIDC			Abdomen
Model	Stat.	Sel.	Gen.	Stat.	Sel.	Gen.	Stat.	Sel.	Gen.
LLaVA-Med	0.00	1.11	3.93	0.00	31.58	33.92	1.16	14.42	36.60
RadFM	46.94	7.78	2.96	0.00	21.05	13.75	10.47	11.54	9.60
Lingshu	7.48	25.56	0.33	20.45	33.83	16.20	37.21	53.85	25.04
Hulu-Med	64.63	27.78	0.88	28.79	30.83	23.87	51.16	52.88	39.89
MedGemma	40.82	24.44	4.55	25.00	21.05	35.94	27.91	45.19	26.11
MedGemma1.5	42.86	23.33	4.41	20.45	27.82	41.93	15.12	49.04	38.18
MedKLIP	0.00	17.78	-	0.00	27.07	-	0.00	28.85	-
KAD	0.00	13.33	-	0.00	16.54	-	0.00	5.77	-
EasyLens	66.67	31.11	5.15	30.30	36.09	45.86	52.33	55.77	40.67

Comparison on ReX, LIDC, and Abdomen. Stat. denotes regional counting, Sel. denotes region selection, and Gen. denotes the lesion-token probability probe defined in Appendix B.2.

Qualitative Results

EasyLens case study on a small pulmonary nodule

Case study of subtle-lesion perception and lesion-aware report generation.

Citation

@article{wang2026easylens, title={EasyLens: A Training-Free Plug-and-Play Subtle-Lesion Representation Amplifier for Medical Vision-Language Models}, author={Wang, Hao and Zeng, Qiwei and Lin, Jinghao and Ye, Shuchang and Yang, Yuezhe and Peng, Yige and Che, Haoyuan and Kim, Jinman and Bi, Lei}, journal={Advances in Neural Information Processing Systems}, year={2026}, url={https://arxiv.org/abs/2606.06379} }