Aims: Integrating multi-stain histopathology data remains a key challenge in digital pathology and cardiovascular research. We developed UNICORN (UNiversal stain Integration network for CORonary classificatioN), a transformer-based deep learning model for multi-stain whole-slide image (WSI) integration. UNICORN enables quantitative classification of coronary atherosclerosis severity, aiming to enhance diagnostic precision, biomarker discovery, and mechanistic understanding in human cardiovascular disease.
Methods: UNICORN is a two-stage, end-to-end trainable transformer architecture that employs domain-specific expert modules for each stain and an aggregation module that learns cross-stain relationships. The model was trained and validated on 1,045 human coronary artery tissue sections from 177 individuals in the Munich Cardiovascular Studies Biobank (MISSION), stained with hematoxylin and eosin (H&E), Elastica van Gieson, von Kossa, and Movat pentachrome. Lesions were classified into five stages according to the American Heart Association and Virmani scheme. Model performance was compared with transformer baselines and human experts using 5-fold cross-validation. Model interpretability was assessed via attention maps highlighting relevant tissue phenotypes.
Results: UNICORN achieved a mean accuracy of 0.68 ± 0.05 and macro F1-score of 0.66 ± 0.04, outperforming baseline models (accuracy 0.64 ± 0.06). The model remained robust to missing stains, with minimal accuracy loss (<6%). Attention visualizations identified biologically and pathologically relevant regions, including fibrous caps, lipid cores, inflammatory infiltrates, and calcifications. UNICORN’s predictions aligned with human expert attention in 70% of slides and surpassed an independent expert’s diagnostic accuracy (0.80 vs. 0.53). The model captured a continuous trajectory of atherosclerosis progression, reflecting underlying disease biology.
Conclusions: UNICORN represents the first deep learning framework to integrate multi-stain histopathology of human coronary arteries, enabling explainable and quantitative classification of atherosclerotic lesion stages. This approach offers translational potential to improve prediction, diagnosis, and biomarker identification in cardiovascular disease, supporting AI-guided precision pathology and the development of individualized treatment strategies for coronary atherosclerosis.
