Deep learning techniques for computer-aided diagnosis in colorectal cancer

Resum

Colorectal cancer (CRC) is the second type of cancer worldwide in terms of mortality and it is the second most common type of cancer in Spain, with 43,370 new cases and 11,131 deaths in 2020. CRC originates from precursor lesions in the colon, known as polyps. Most of polyps are benignant, although some of them may evolve into CRC depending on their histological characterization and therefore it is important to detect them as soon as possible. The standard method for finding and treating colon polyps is colonoscopy, a medical procedure where an endoscopist observes and carefully examines the colonic mucosa looking for potential lesions. The main goal of the PhD thesis presented here is the creation of Deep Learning models, specifically Convolutional Neural Networks, for the detection and classification of colorectal polyps in real time during colonoscopy. To achieve this goal, two specific objectives were defined: (i) creating a detection model able to detect the presence of polyps in colonoscopy images (detection), as well as providing their exact location (localization); (ii) creating a classification model able to predict the histological characterization of a polyp previously identified by the detection model. This research work presents the development and validation process of these models, complemented with additional activities such as the development of a public database of colorectal polyp images and videos. The creation of these models enables the development of a CAD system that assists endoscopists in real time during colonoscopy, acting as a second observer to improve the identification of polyps and their optical diagnosis. A greater number of polyps could be detected with the use of this system, therefore allowing an earlier treatment and lowering CRC incidence. Also, the number of resected polyps sent to histological analysis could be reduced, reducing costs and improving the efficiency of the healthcare system.