Show abstract - PDF - Reviews

Fully-supervised machine learning has been established as an effective method for medical image segmentation. However, it requires large amounts of expert-annotated data, which can be a bottleneck in certain applications. Unsupervised methods like anomaly localization have proven their potential without relying on any labeled data, making them potentially much more scalable than fully supervised methods. Despite their scalability advantages, unsupervised and self-supervised methods have not yet fully reached the performance level of fully supervised models. As a first step to close this gap, we propose an approach that combines both concepts. We fine-tune a pre-trained anomaly localization model, namely a self-supervised denoising auto-encoder, using varying amounts of labeled training data in a supervised manner. Overall this approach exhibits superior performance compared to a model trained from scratch, especially in a low labeled training data regime.
Hide abstract