Abstract
Glaucoma is a chronic, progressive eye disease projected to affect 111 million people by 2040, which has motivated numerous automatic detection methods, most of them relying on black-box models with limited interpretability. We introduce an inherently interpretable framework for glaucoma screening that extracts 20 clinically relevant concepts: cup-to-disc ratios, ISNT-sector areas, shape descriptors, and regional color statistics, from color fundus images and feeds them to a Neural Additive Model in which each concept is handled by a dedicated subnetwork. Despite containing only 2,587 trainable parameters, the model rivals far larger deep neural networks: on the Rim-One DL benchmark, it reaches an accuracy of 0.901, an F1 score of 0.855, and an AUROC of 0.945 on the by-Random split, and remains competitive under domain shift on the by-Hospital split (accuracy 0.851, F1 0.774, AUROC 0.894), while outperforming logistic-regression and gradient-boosting baselines trained on the same concepts. Crucially, every prediction is accompanied by intrinsic, per-concept local explanations that remain clinically meaningful, obtained without any post hoc explanation method, thereby supporting the transparency requirements placed on medical AI and contributing to greater trust in AI systems in medicine.
Paper Type
Full Paper
DOI
10.62036/ISD.2026.82
Neural Additive Model based framework for interpretable glaucoma screening
Glaucoma is a chronic, progressive eye disease projected to affect 111 million people by 2040, which has motivated numerous automatic detection methods, most of them relying on black-box models with limited interpretability. We introduce an inherently interpretable framework for glaucoma screening that extracts 20 clinically relevant concepts: cup-to-disc ratios, ISNT-sector areas, shape descriptors, and regional color statistics, from color fundus images and feeds them to a Neural Additive Model in which each concept is handled by a dedicated subnetwork. Despite containing only 2,587 trainable parameters, the model rivals far larger deep neural networks: on the Rim-One DL benchmark, it reaches an accuracy of 0.901, an F1 score of 0.855, and an AUROC of 0.945 on the by-Random split, and remains competitive under domain shift on the by-Hospital split (accuracy 0.851, F1 0.774, AUROC 0.894), while outperforming logistic-regression and gradient-boosting baselines trained on the same concepts. Crucially, every prediction is accompanied by intrinsic, per-concept local explanations that remain clinically meaningful, obtained without any post hoc explanation method, thereby supporting the transparency requirements placed on medical AI and contributing to greater trust in AI systems in medicine.
Recommended Citation
Potrzebowski, P. & Czyżewski, A.(2026). Neural Additive Model based framework for interpretable glaucoma screening. In M. Valenta, B. Mannová, R. Pergl, A. Przybylek, M. Lang, H. Linger, C. Schneider, N. Iivari, & E. Insfran (Eds.), Making ISD Sustainable: Reloaded with AI and Automation (ISD2026 Proceedings). Prague, Czech Republic: Czech Technical University in Prague. ISBN: 978-80-01-07585-2. https://doi.org/10.62036/ISD.2026.82