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

Over the last decades, medical imaging techniques haveplayed a crucial role in healthcare, supporting radiologistsand facilitating patient diagnosis. With the advent of fasterand higher-quality imaging technologies, the amount of datathat is possible to collect for each patient is paving the waytoward personalised medicine. As a result, automating simpleimage analysis operations, such as lesion localisation andquantification, would greatly help clinicians focus energyand attention on tasks best done by human intelligence.Most recently, Artificial Intelligence (AI) research is acceleratingin healthcare, providing tools that often perform on par oreven better than humans in conceptually simple image processingoperations. In our work, we pay special attention tothe problem of automating semantic segmentation, where animage is partitioned into multiple semantically meaningfulregions, separating the anatomical components of interest.Unfortunately, developing effective AI segmentation tools usuallyneeds large quantities of annotated data. Conversely,obtaining large-scale annotated datasets is difficult in medicalimaging, as it requires experts and is time-consuming.For this reason, we develop automated methods to reduce theneed for collecting high-quality annotated data, both in termsof the number and type of required annotations. We makethis possible by constraining the data representation learnedby our method to be semantic or by regularising the modelpredictions to satisfy data-driven spatio-temporal priors. Inthe thesis, we also open new avenues for future research usingAI with limited annotations, which we believe is key todeveloping robust AI models for medical image analysis.