Semi-supervised and weakly-supervised learning with spatio-temporal priors in medical image segmentation

Over the last decades, medical imaging techniques have played a crucial role in healthcare, supporting radiologists and facilitating patient diagnosis. With the advent of faster and higher-quality imaging technologies, the amount of data that is possible to collect for each patient is paving the way toward personalised medicine. As a result, automating simple image analysis operations, such as lesion localisation and quantification, would greatly help clinicians focus energy and attention on tasks best done by human intelligence. Most recently, Artificial Intelligence (AI) research is accelerating in healthcare, providing tools that often perform on par or even better than humans in conceptually simple image processing operations. In our work, we pay special attention to the problem of automating semantic segmentation, where an image is partitioned into multiple semantically meaningful regions, separating the anatomical components of interest. Unfortunately, developing effective AI segmentation tools usually needs large quantities of annotated data. Conversely, obtaining large-scale annotated datasets is difficult in medical imaging, as it requires experts and is time-consuming. For this reason, we develop automated methods to reduce the need for collecting high-quality annotated data, both in terms of the number and type of required annotations. We make this possible by constraining the data representation learned by our method to be semantic or by regularising the model predictions to satisfy data-driven spatio-temporal priors. In the thesis, we also open new avenues for future research using AI with limited annotations, which we believe is key to developing robust AI models for medical image analysis.