Colorectal Polyp Classification using Convolutional Neural Networks

Samenvatting

Colorectal Polyps (CRPs) are precursor lesions for Colorectal Cancer (CRC). The survival rate of CRC is heavily correlated with the stage of diagnosis, therefore, early detection and removal of CRPs that bear risk of progression into cancer is essential. The current medical protocol dictates that all CRPs encountered during colonoscopy procedures should be resected and undergo histopathological evaluation, with inherent
costs and risk of complications. To prevent from unnecessary removal of polyps and enable adoption of new medical protocols, Computer-aided diagnosis (CADx) systems are desired to assist clinicians with characterization of CRPs during colonoscopy.
In this study, several approaches are explored to improve reliability, robustness and performance of a Convolutional Neural Network (CNN) for classification of CRPs. Firstly, the influence of the image content in the region of interest (ROI) provided to the CNN, on the classification performance is investigated. Furthermore, the added value of a contrastenhancement algorithm and domain-specific pre-training stage on classification performance are examined. Moreover, several model confidence calibration methods are explored, in order to obtain a model that is well-calibrated, thereby increasing the reliability and trustworthiness of classification confidence scores. Lastly, an User Interface has been developed for testing the feasibility of the proposed CADx system in real-time clinical colonoscopy procedures.

Datum prijs	24 aug 2021
Originele taal	Engels
Begeleider	Fons van der Sommen (Afstudeerdocent 1)