Deep Learning models offer flexible solutions which can be applied to solve a broad range of complex tasks based on their ability to extract even subtle patterns present in data. Their ability to swiftly process large datasets and deliver highly accurate results has led to them being increasingly harnessed in the medical field to support diagnosis of diseases or conditions such as colorectal cancer. 

Colorectal cancer has a high mortality rate which can be decreased by early detection and intervention. Patients with Inflammatory Bowel Disease are at heightened risk of developing colorectal cancer and thus undergo regular surveillance, but recognition of premalignant changes is difficult and inaccurate due to inflammation. These issues can further be exacerbated based on colonoscopy operator differences, such that many changes are missed. 

Recognition of pre-cancerous lesions can reasonably be improved by deep learning techniques by highlighting suspicious lesions for endoscopists during colonoscopy screenings, which is why this thesis investigated various approaches to detect and classify potential precancerous lesions during surveillance colonoscopy in patients with Inflammatory Bowel Disease. Previous research in the field was analyzed to identify and present the current state of the art. Further, a Deep Learning pipeline was designed to overcome limitations in available training data quality and quantity, and several model architectures were applied to the task. Lastly, the overall performance of the proposed Deep Learning system was evaluated. 

The literature research returned a small number of initial studies into the area of pre-cancerous learning detection in patients with Inflammatory Bowel Disease, with promising results achieved when utilizing a training dataset specifically constructed of lesions found in patients with the disease. In contrast, when utilizing conventional computer-assisted diagnosis tools trained on lesions found in a general population, no improvements were reported. 

Although the classification performance achieved by the subsequently constructed Deep Learning pipeline was not sufficient for clinical usage, this thesis project demonstrated the importance of utilizing subsets for training and testing which are split on a per-lesion basis to ensure the ability of the Deep Learning model to generalize its training onto previously unseen data. When utilizing pretraining datasets derived from a general population, classification accuracy could be improved, and this effect was shown to be able to be combined with utilization of ImageNet weights. Impact was shown to be determined based on whether pretraining and training datasets had overlapping classes. Further, comparison of models yielded highest accuracy when utilizing an InceptionNet-v4 architecture, while both a CSPDarknet-53 and a DenseNet-201 architecture achieved a balance between significantly reduced training duration and preserved accuracy. Lastly, the application of conformal risk control was presented for usage in a pre-cancerous lesion classification context, resulting in more information about the Deep Learning model's classification confidence being available, and thereby potentially reducing the risk of mistreating malignant lesions as benign when identified during surveillance colonoscopy.