Semantic Robustness of Deep Neural Networks in Ophthalmology: A Case Study with Colour Fundus Imaging

ICLR 2026 Conference Submission4844 Authors

Keywords: Semantic Robustness, Retinal Image, DIRECT optimisation

Abstract:

Despite the success of Deep Neural Networks (DNNs) in ophthalmic tasks, their robustness in real-world clinical settings remains uncertain.
This paper presents a case study on the semantic robustness of DNN models for retinal imaging. We first introduce a novel optimisation algorithm, DIRECT-LSR, to identify worst-case robustness against clinically relevant semantic perturbations, including geometric transformation, illumination distortion, and motion blur. Our method provides a reliable lower bound with theoretical guarantees, enabling a practical black-box robustness validation approach. Evaluating various commonly used DNN models on colour fundus photograph datasets, we demonstrate their vulnerability to semantic perturbations, particularly geometric transformations that drastically reduce model accuracy despite preserving clinically relevant features. As a secondary contribution, we show that a randomised data augmentation strategy can serve as an effective and accessible defence mechanism to improve models' reliability. However, since the performance gaps between clean and perturbed images persist, our results also highlight the need for more advanced defences in future work, offering insights for developing more reliable artificial intelligence systems.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 4844