Deep learning for the detection of colon polyps with malignant potential: ex vivo classification using feature-enhanced optical coherence tomography (OCT) images.

Abstract

Colorectal cancer (CRC) is the second leading cause of cancer-related morbidity and mortality in both men and women globally. CRC predominantly arises from dysplastic polyps that, over time, progressively evolve into malignancies. Population-wide screening through colonoscopy remains the cornerstone of CRC prevention. Optical coherence tomography (OCT) has the potential to increase the effectiveness and reduce the cost associated with colonoscopic screening. However, conclusive evidence that OCT can effectively detect pre-cancerous changes is still lacking. This study introduces a novel framework to address this challenge by extracting additional features, which can serve as biomarkers of disease, from ex vivo OCT images of colon polyps. These include first and second-order intensity and fractal statistics, as well as spectral characteristics and scatterer size, which depend on sub-cellular and biochemical tissue variations. Feature-enhanced images derived from these biomarkers were combined with intensity images and integrated into a deep-learning classification model decision-level fusion. This approach achieved 88.3% accuracy, 93.5% sensitivity, 77.9% specificity, and an AUC of 0.857 in distinguishing benign (normal and hyperplastic) polyps from cases with malignant potential (adenoma and sessile serrated adenoma), demonstrating the potential of this novel approach to enhance the role of OCT in improving CRC screening outcomes.