On the Generalizability of Deep Learning-based Medical Image Segmentation Methods

Abstract

Despite achieving state-of-the-art performance in lab-conditions, deep learning-based systems often exhibit significant performance degradation when deployed in practical settings. This is referred to as generalization failure. Why and how this occurs has only recently started to be understood, and there has consequently been limited research towards developing generalizable methods for deep learning. This thesis attempts to address generalization failure in the domain of medical image segmentation, in particular on the polyp segmentation task. Recent analyses of generalizability is discussed, which is then used to inform the development of a number of novel methods. This includes a simple dual-decoder architecture, an augmentation strategy which incorporates a generative polyp inpainter, a training method referred to as Consistency Training, and finally, several ensemble models for which the constituent predictors are trained using Consistency Training. These methods are then evaluated through multiple quantitative studies. As the extent to which methods used as baselines in this thesis affect generalization is not particularly well understood, this thesis also contributes a quantitative analysis of the the impact of the choice of model architecture, data augmentation, and ensemble-models on generalization. The results show that Consistency Training facilitates increased generalitation over data augmentation. The use of the inpainter as a component of data augmentation, however, limits the possible improvements com- pared to regular augmentation. Ensembles improve generalization, albeit to a somewhat lesser extent than the aforementioned methods. Finally, the choice of model architecture, including the use of a secondary decoder, is shown to have negligible effects on generalization. These results were all explored with respect to theory presented in other literature. These findings are then analyzed and used to inform a number of hypotheses which are suggested as points of further study. Several improvements to the proposed methods were also suggested, in particular with regards to Consistency Training, which shows significant promise towards further mitigating generalization failure.