KLASIFIKACIJA KOLOREKTALNIH ADENOMA ZASNOVANA NA DUBOKOM UČENJU – MODEL ZASNOVAN NA PODACIMA ZA EFIKASNOST RESURSA

Abstract

We evaluated the performance and practical impact of several deep learning models, including ResNet and transformer-based architectures, for classifying colorectal adenomas with and without high-grade dysplasia in histopathology images, to support an artificial intelligence-assisted triage workflow that reduces pathologist workload while maintaining diagnostic quality and workflow efficiency. Models were trained and validated on 1,669 histopathology image tiles from colorectal polyp samples (368 with high-grade dysplasia, 767 low-grade adenomas, and 534 non-neoplastic controls) that were expert-reviewed and analysed at 100× magnification. Performance was assessed using 5-fold cross-validation, and Grad-CAM++ heat maps were generated to visualise model attention and support explainable pathologist review. ResNet101 achieved 99.17% accuracy in the best fold and 97.40% on average. The explanatory heat maps consistently high lighted diagnostically relevant epithelial regions, supporting rapid pathologist verification. In a simulated workflow of 1,000 colorectal polyps, artificial intelligence-assisted triage reduced pathologist review volume by 41–61% while maintaining 98.91% detection accuracy for high-grade dysplasia and 99.47% accuracy for adenomas without high-grade dysplasia. These reductions correspond to meaningful full-time equivalent savings in high-volume gastrointestinal pathology practice, where repetitive biopsy assessment contributes substantially to the daily workload. These findings indicate that deep learning models can accurately classify colorectal adenomas and detect high grade dysplasia, and that artificial intelligence-assisted triage and sign-out, supported by interpretable heat maps, may provide a practical pathway towards more sustainable and efficient dig ital pathology workflows.