Semi-Supervised Learning Approach for Bladder Cancer Diagnosis

Recent studies have made great strides in reducing the labeling burden in deep learning algorithms by requiring that only a subset of the dataset be labeled. These are called semi-supervised learning algorithms (SSL). In this thesis we explore a type of semi-supervised learning algorithm that focuses on a technique called consistency regularization and self-ensembling to leverage the unlabeled portion of the dataset. Consistency regularization has been used as a driving force in SSL to great success in datasets of natural images, but it has not been used in image pathology where the dataset consists of cell patterns. We successfully implement an SSL algorithm based on the VGG-16 neural network, which combines techniques introduced by state-of-the-art algorithms such as FixMatch and the Π model. Our study shows that our algorithm can leverage the unlabeled portion of a dataset to improve a model’s accuracy by up to 19% from the baseline.