BioML at CVPR 2026: Exploring Robust AI for Medical Imaging

8 Juni 2026

Dr. Athanasios Angelakis represented the BioML Lab at the Research Institute CODE, Universität der Bundeswehr München, at the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2026), held in Denver, Colorado, from 3 to 7 June 2026.

At the PHAROS AI Factory for Medical Imaging & Healthcare (PHAROS-AIF-MIH) workshop on 3 June, he presented the paper “Extending ZACH-ViT to Robust Medical Imaging: Corruption and Adversarial Stress Testing in Low-Data Regimes,” co-authored with Prof. Dr. Marta Gomez-Barrero. The workshop brought together researchers working on trustworthy AI for medical imaging and healthcare.

The study investigates how ZACH-ViT, a compact vision transformer, performs when medical images are affected by noise, blur, compression and other distortions. The evaluation covered seven medical-imaging datasets, with models trained from scratch using only 50 images per class. Among the four compact models evaluated, ZACH-ViT achieved the best average ranking across datasets on both unaltered images and images subjected to common degradations. Deliberately crafted adversarial perturbations, however, substantially reduced performance across all models, highlighting an ongoing research challenge.

The findings underline the value of evaluating medical AI beyond performance on clean test images and contribute to research on compact models that can learn from limited data while accounting for variations in image quality.

The paper is published in the **CVPR 2026 Workshops proceedings** and is available through CVF Open Access.