Histopathological Assessment of Myocardial Ischemia-Reperfusion Injury Using Transformer-Based Artificial...

Myocardial ischemia-reperfusion injury (MIRI) poses diagnostic challenges due to complex histopathological changes.

This study aimed to develop an intelligent framework for evaluating MIRI on hematoxylin-eosin-stained slides, to compare major deep learning architectures, and to determine the advantages of transformer models across multiple interventions and time points.

A total of 1280 whole-slide images (~62,000 tiles) from public datasets were analyzed across antioxidant, β-blocker, calcium channel blocker, and control groups at 6, 24, and 72 hours. Seven model families (convolutional neural networks, recurrent neural networks, long short-term memory networks, autoencoders, graph convolutional networks, variational autoencoders, and transformers) were trained under unified preprocessing, with generative adversarial networks used exclusively for leakage-free augmentation. Weak supervision used a clustering-constrained attention multiple-instance learning strategy, and segmentation applied a Transformer-UNet. Data were split into 8:1:1 at the subject level, with 5-fold cross-validation.

The transformer achieved the best performance (accuracy=0.942; area under the curve=0.982; and F1-score=0.958). Segmentation Dice scores were 0.847 (necrosis) and 0.821 (apoptosis). Predictions strongly agreed with expert measurements (r=0.886; Bland-Altman limits +5% or -5%), and attention maps aligned with necrotic borders and inflammatory foci. Temporal trends matched biological expectations, with the antioxidant group showing the most stable improvement.

Transformer-based pathology offers accurate, robust, and interpretable assessment of MIRI and provides a scalable framework for dynamic injury quantification and therapeutic evaluation.