Recent advancements in deep learning-based AI technologies provide an automatic tool to revolutionise medical image computing. Training a deep learning model requires a large amount of labelled data. Acquiring labels for medical images is extremely challenging due to the high cost in terms of both money and time, especially for the pixel-wise segmentation task of volumetric medical scans. However, obtaining unlabelled medical scans is relatively easier compared to acquiring labels for those images.

This work addresses the pervasive issue of limited labels in training deep learning models for medical imaging. It begins by exploring different strategies of entropy regularisation in the joint training of labelled and unlabelled data to reduce the time and cost associated with manual labelling for medical image segmentation. Of particular interest are consistency regularisation and pseudo labelling. Specifically, this work proposes a well-calibrated semi-supervised segmentation framework that utilises consistency regularisation on different morphological feature perturbations, representing a significant step towards safer AI in medical imaging. Furthermore, it reformulates pseudo labelling in semi-supervised learning as an Expectation-Maximisation framework. Building upon this new formulation, the work explains the empirical successes of pseudo labelling and introduces a generalisation of the technique, accompanied by variational inference to learn its true posterior distribution. The applications of pseudo labelling in segmentation tasks are also presented. Lastly, this work explores unsupervised deep learning for parameter estimation of diffusion MRI signals, employing a hierarchical variational clustering framework and representation learning.