Development of Image Processing Algorithms for the AutomaticScreening of Colon Cancer

Abstract

Colorectal cancer (CRC) is one of the most commonly diagnosed cancers among both genders and its incidence rate is continuously increasing. CRC starts from small non-cancerous growths of tissue on the wall of the colon (large bowel) or rectum. Most polyps are harmless, but some can develop into CRC over time. Currently, colonoscopy is the golden standard method for the detection and removal of precancerous polyps. Colonoscopy, however, is an operator-dependent procedure and requires skilled endoscopists. Studies have shown that the polyp miss rate is around 25\% for certain cases. This miss rate has drawn the attention of engineers and computer scientists, including our group, for decades to develop a computer-aided polyp detection system that can help clinicians reduce this polyp miss rate during colonoscopy. This thesis has primarily contributed towards the investigation of the difficulties and challenges to develop an accurate automatic polyp detection and segmentation using deep learning approaches. Experimental results showed that deep learning is a promising approach to computerize colon polyp detection and segmentation, and it offers various approaches to improve the overall performance of the detection. In general, a massive amount of training data is the key to achieve desirable performance as there are already excellent CNN-based feature extractors. However, there is a lack of available training data, and manual polyp labeling of video frames is difficult and time-consuming. We showed that deep learning can be used to semi-automatically annotate video frames and produce 96\% of the Dice similarity score between the polyp masks provided by clinicians and the masks generated by our framework. We also showed that conditional GAN (CGAN) could be used to generate synthetic polyps to enlarge the training samples and improve the performance. The results demonstrated that deep learning-based models are vulnerable to small perturbations and noises. We found out that the bidirectional temporal information is essential to make CNN-based detection more reliable and less vulnerable.