Methodological comparative study of machine learning techniques for the detection of polyps in endoscopy images

This thesis investigates the problem of automatic polyp segmentation in gastrointestinal endoscopy using machine learning semantic segmentation approaches.In our study, we compare six different architectures that are common in medical imaging: UNet, AttentionUnet, SegResNet, EffiSegNet, UNETR, and SwinUNETR.We used PolypGen dataset's centres C1-C5 to train all compared models from scratch, and centre C6 for evaluation.A notable challenge in our study is the significant domain shift observed in every center, and especially the evaluation center C6, which includes different populations, lighting conditions, image resolutions and field of view.The pipeline is created using Pytorch and Lightning frameworks for both the model architectures and the training loops.Hydra framework is used for configuration management, and MONAI for medical imaging components.Results are logged and visualized in TensorBoard.Our pipeline evaluates the six beforementioned architectures, using a variety of metrics common in segmentation tasks, and then creates an ensemble method, that combines predictions from the three best performing architectures through a soft voting mechanism.The ensemble method seems to outperform all previously tested individual architectures, by a slight margin.