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

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Cardiovascular research

Background:

• Myocardial ischemia-reperfusion injury (MIRI) presents diagnostic challenges due to complex histopathological changes.
• Accurate assessment of MIRI is crucial for patient outcomes and therapeutic development.

Purpose of the Study:

• To develop an intelligent framework for evaluating MIRI using hematoxylin-eosin-stained slides.
• To compare the performance of various deep learning architectures for MIRI assessment.
• To determine the advantages of transformer models in analyzing MIRI across different interventions and time points.

Main Methods:

• Analysis of 1280 whole-slide images from public datasets across antioxidant, β-blocker, calcium channel blocker, and control groups at 6, 24, and 72 hours.
• Training and comparison of seven deep learning model families, including transformers, using unified preprocessing and weak supervision with attention multiple-instance learning.
• Segmentation of MIRI using a Transformer-UNet architecture, with data split and 5-fold cross-validation.

Main Results:

• Transformer models achieved superior performance with an accuracy of 0.942, AUC of 0.982, and F1-score of 0.958.
• High segmentation Dice scores for necrosis (0.847) and apoptosis (0.821) were obtained.
• Model predictions showed strong agreement with expert measurements (r=0.886), and attention maps highlighted key pathological features.

Conclusions:

• Transformer-based pathology provides an accurate, robust, and interpretable method for assessing MIRI.
• The developed framework enables scalable, dynamic quantification of injury and evaluation of therapeutic efficacy.
• This approach holds promise for advancing MIRI research and clinical diagnostics.