Deep learning classification of colon polyps for colorectal cancer detection (CRC) using feature-enhanced ex vivo optical coherence tomography (OCT) images

Colorectal cancer (CRC) is a leading cause of cancer deaths, with thousands of new cases each year worldwide. CRCs mainly develop from dysplastic polyps, which gradually progress into malignancies. Population-based screening with colonoscopy is a crucial cancer prevention strategy, potentially decreasing CRC mortality by up to 57%. Despite its effectiveness, colonoscopy can be improved. Recent efforts to enhance colonoscopic efficiency tried to reduce the procedure time and cost by leaving diminutive polyps (<5mm) in place. However, such a strategy introduces additional risks since it implies leaving 80% of the, now-resected, polyps in place. Thus, technological advancements are needed to minimize the risk from unresected polyps. OCT can aid in detecting and classifying colon polyps, but its application in colon imaging lags behind other clinical uses due to its dependence on images of tissue microstructure without any biochemical information, necessary, to distinguish early signs of cancer. This study proposes a novel approach to overcome limitations in interpreting OCT images for CRC screening. Ex vivo OCT images of colon polyps were used to extract features that remain unused and indicate sub-resolution and biochemical tissue changes, such as scatterer size and group velocity dispersion. Along with the intensity images, feature images were created and used for deep learning (DL) classification with late fusion to distinguish benign (normal and hyperplastic) from malignant (adenoma and SSA) polyps. The novel DL method resulted in 88.3 % correct classification rate per image.