Scholarlink

Pre-trained foundation models have advanced automated diabetic retinopathy (DR) screening by achieving high diagnostic accuracy from fundus images. To promote safe clinical deployment, uncertainty estimation is often integrated during fine-tuning to measure model confidence and detect out-of-distribution (OOD) inputs. Among challenging OOD cases, adversarial examples represent a subtle form of boundary OOD, where perturbations remain visually plausible but induce confident misclassifications. In this study, we systematically examine how uncertainty-aware fine-tuning (Uncertain-FT) affects model robustness against boundary OOD attacks on DR grading using the APTOS2019 benchmark. Adversarial examples are generated using Projected Gradient Descent (PGD) and Carlini & Wagner (C&W) attacks under both white-box and transfer scenarios, and evaluated on two representative pre-trained models, RetFound and RetiZero. Results show that Uncertain-FT offers limited improvements in adversarial robustness, with attack success rates remaining high under both PGD and C&W attacks. Moreover, many adversarial examples still yield low uncertainty scores, escaping detection, whereas human experts remain unaffected. These findings reveal critical limitations in current uncertainty quantification for detecting boundary OOD threats in DR screening and underscore the need for more robust and trustworthy uncertainty modeling in ophthalmic AI.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

On the Limits of Uncertainty-Aware Fine-Tuning for Robust Diabetic Retinopathy Screening

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ting Xu,
  • Jie Zhang,
  • Meng Wang,
  • Hongyu He,
  • Xinbao Zou

摘要

Pre-trained foundation models have advanced automated diabetic retinopathy (DR) screening by achieving high diagnostic accuracy from fundus images. To promote safe clinical deployment, uncertainty estimation is often integrated during fine-tuning to measure model confidence and detect out-of-distribution (OOD) inputs. Among challenging OOD cases, adversarial examples represent a subtle form of boundary OOD, where perturbations remain visually plausible but induce confident misclassifications. In this study, we systematically examine how uncertainty-aware fine-tuning (Uncertain-FT) affects model robustness against boundary OOD attacks on DR grading using the APTOS2019 benchmark. Adversarial examples are generated using Projected Gradient Descent (PGD) and Carlini & Wagner (C&W) attacks under both white-box and transfer scenarios, and evaluated on two representative pre-trained models, RetFound and RetiZero. Results show that Uncertain-FT offers limited improvements in adversarial robustness, with attack success rates remaining high under both PGD and C&W attacks. Moreover, many adversarial examples still yield low uncertainty scores, escaping detection, whereas human experts remain unaffected. These findings reveal critical limitations in current uncertainty quantification for detecting boundary OOD threats in DR screening and underscore the need for more robust and trustworthy uncertainty modeling in ophthalmic AI.