Prediction Pathological Structure of Computed Tomographic Colonography Polyps via Machine Learning

Computed tomographic colonography (CTC) has shown to be an effective means of detecting the precursor polyps leading to colorectal cancer (CRC). Accurately diagnosing these polyps can assist the overall treatment plan and follow-up examination determination to prevent CRC. In this work, we examine a machine learning approach to predicting the three main pathological structure type of polyps, comparing hyperplastic and serrated adenomas (HS), tubular adenomas (TA), and advanced neoplasms (AN). We find that imaging textures can differentiate any two pathological groups with AUC scores ranging from 0.624-0.749 for polyps less than 10mm in size and 0.731-0.887 for polyps greater than 10mm. Classification of all three pathology groups simultaneously shows a sensitivity of 61.6% for identifying the highest risk group of advanced neoplasms.