Occurrence-Aware Prototype Graph Learning for Interpretable Multi-Label Disease Diagnosis
IEEE Access · 2026 · 14 · 135783-135802
DOI: 10.1109/access.2026.3728910 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Mansoor Hayat und 1 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 52 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Thailand Science Research and Innovation Fund, Chulalongkorn University.
Abstract (Originalsprache)
Prototype-based learning is attractive for medical image diagnosis because predictions can be linked to visually inspectable training patterns. In multi-label disease diagnosis, however, prototype evidence is easily confounded by co-occurring pathologies and by background anatomy that dominates the feature map. This paper proposes Occurrence-Aware Prototype Graph Learning (OAPG), an interpretable multi-label framework that extends cross- and intra-image prototypical learning with two complementary modules. First, an Occurrence-Aware Gating (OAG) module learns a spatial evidence map and applies a soft offset gate before prototype matching, upweighting pathologically relevant regions while preventing complete suppression of contextual anatomy. Second, a Prototype Graph Network (PGN) performs lightweight relational refinement over prototype activation scores through a learnable inter-prototype adjacency matrix. The original prototype distances and similarity heatmaps remain available for case-based visual explanation, while the graph-refined activations improve final multi-label prediction. Experiments on NIH ChestX-ray14 and ODIR-5K show consistent improvements over the strongest prototype-based baseline. On ChestX-ray14, OAPG achieves a mean AUC of 0.836, improving over CIPL by 0.8 percentage points in mean AUC and 0.9 percentage points in mean F1. On ODIR-5K, OAPG obtains mean AUC values of 0.892 and 0.912 on the on-site and off-site test sets, respectively, exceeding CIPL by 1.2 and 0.9 percentage points. Paired bootstrap testing confirms that these benchmark-level improvements are statistically significant, and complexity analysis shows that the added modules introduce only modest computational overhead.
Förderung
- Thailand Science Research and Innovation Fund, Chulalongkorn University
Zitieren
APA
Hayat, M., & Aramvith, S. (2026). Occurrence-Aware Prototype Graph Learning for Interpretable Multi-Label Disease Diagnosis. IEEE Access, 14, 135783-135802. https://doi.org/10.1109/access.2026.3728910
Vancouver
Hayat M, Aramvith S. Occurrence-Aware Prototype Graph Learning for Interpretable Multi-Label Disease Diagnosis. IEEE Access. 2026;14:135783-135802. doi:10.1109/access.2026.3728910
BibTeX
@article{Hayat2026,
author = {Hayat, Mansoor and Aramvith, Supavadee},
title = {Occurrence-Aware Prototype Graph Learning for Interpretable Multi-Label Disease Diagnosis},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {135783--135802},
doi = {10.1109/access.2026.3728910}
}
EUVIMED