Deep Learning in Precancerous Lesions Detection during Surveillance Colonoscopies of IBD Patients.

Abstract

Deep Learning (DL) models have developed tremendously over the last couple of decades in their ability to train across large datasets and give fast and accurate results across a varied number of tasks like image classification and segmentation. This is the reason why DL models are being increasingly adopted for aiding medical professionals in the diagnosis and detection of various medical conditions like colorectal cancer (CRC). There happens to be a specific group of patients who suffer from Inflammatory Bowel Disease (IBD) who are at a significantly higher risk of developing CRC, which is why they undergo surveillance colonoscopies. However, precancerous lesions that have the potential of developing into cancer can sometimes be difficult to spot, identify and observe changes in, during colonoscopies. IBD patients can have internal scarring of tissue, which makes it even more difficult to detect precancerous lesions and spot the changes in them during surveillance colonoscopies. Modern DL models can be useful in aiding the detection and identification of these precancerous lesions, which is why in this thesis, various DL-based approaches for the detection and clustering of precancerous lesions during surveillance colonoscopies of IBD patients were tested. Both a supervised object-detection-based approach on a labelled dataset, and an unsupervised image-clustering-based approach were tried out using pre-designed DL models. Furthermore, it was investigated whether the colour channel separation and possible recombination of certain colour channels of the images in the dataset could help improve the detection of precancerous lesions, and make the object detection model more accurate.

Some results of the unsupervised image-clustering-based approach looked promising, but it was unable to segregate each type of potential precancerous findings into separate clusters. The supervised learning-based approach that did object detection worked very well with the labelled dataset used in this project. The colour channel separation and recombination of images in the dataset gave a significant improvement to the performance of the object detection model, particularly when the images in the dataset consisted of only the blue channel of the original RGB images.