Robustness of medical imaging models with physics-based image quality degradations

Tyler Ward, Abdullah Imran

Keywords: Data augmentation, Deep learning, Dermoscopy, Image quality, Medical physics, X-ray

Abstract: Variability in the quality of medical images poses a significant challenge to the reliability of deep learning models employed for medical image analysis and computer-aided diagnostic tasks. Ideally, segmentation and classification networks would be trained on clean, well-curated datasets, but oftentimes real clinical data exhibits degradations caused by suboptimal acquisition conditions. To investigate the robustness of popular medical imaging models under such conditions, we synthetically corrupt publicly available X-ray and dermoscopic datasets with degradations based on real-world medical physics, and develop a comprehensive evaluation framework with multiple segmentation and classification architectures. In addition, we introduce a corruption-level prediction objective to quantify models’ ability to infer image quality directly from degraded inputs. Based on the models' performance across these tasks, we analyze how different corruption levels influence performance and how exposure to physics-informed degradations affect generalizability. Our results demonstrate that models trained with realistic corruption processes exhibit improved robustness in both segmentation and classification accuracy, while also achieving reliable prediction of corruption severity.

Primary Subject Area: Detection and Diagnosis

Secondary Subject Area: Segmentation

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 214