Semantic Segmentation and Classification of OCT Colorectal Polyp Images

Colorectal cancer, being the second leading cause of cancer-related deaths, necessitates effective population screening methods, such as colonoscopy. Traditional histopathologic analysis of biopsies obtained during colonoscopy, although effective, is both costly and time-consuming, incurring a significant burden on the healthcare system. One approach to reduce this load is to “leave in situ” small benign masses and remove the ones with malignant potential. To achieve this goal, a highly accurate, in vivo, evaluation of the polyps is required. Here, a novel methodology is proposed, utilizing deep learning algorithms, to segment and classify polyps from en face ex vivo Optical coherence tomography (OCT) images. A DC-UNet based model is used to classify samples as either benign (normal and hyperplastic) or malignant potential (adenoma and sessile serrated adenoma) samples. The dataset consisted of 143 en face OCT images which were annotated by a histologist to create masks of the ground truth. Image segmentation was then performed using DC-UNet. The algorithm was also trained to classify the images. The training accuracy was 98.08%. Additionally, the model achieved a dice coefficient (F1 score) of 0.5923, Jaccard score of 0.4234, prediction accuracy of 72%, sensitivity 72%, and specificity 66%. These preliminary results indicate that OCT has the potential to increase the accuracy of early detection of colon malignancies. In addition, with further improvements in the future, including a larger in vivo study, OCT can also become the enabler of a leave-in situ approach that can significantly reduce the burden of colonoscopies on the healthcare system.