Cardiac Image Segmentation: towards better reliability and generalization.

Cardiac image segmentation is the problem of learning the anatomical semantics of each voxel in a three-dimensional heart image. In clinical practice, radiologists are delegated to draw contours manually, encompassing the structures of interest. The process is lengthy, monotonous, and prone to subjective errors. Starting from the 1970s, researchers have thoroughly investigated the possibility of automating this task. Automated CMR segmentation can help clinicians interpreting the medical conditions, speeding up diagnoses, increasing monitoring reliability, facilitating surgical planning, and enabling vast population studies. Overall, it would make a strong contribution to the battle against cardiovascular diseases (CVDs), estimated to cost 31\% of all global deaths. During the last decade, this automation attempt has been lead by deep learning. Between 2013 and 2015, deep learning techniques became popular, and more and more papers on the topic went public. When the MICCAI conference of 2017 hosted the ACDC Challenge, nine participants out of ten implemented a deep convolutional architecture to fulfill the segmentation task. This brief time window represents a drastic change in the field. Results reveal that deep learning methods can successfully classify patient data and get highly accurate segmentation results. However, these approaches require fully annotated datasets, which must capture the anatomical variability of heart images. Collecting so much data requires extensive human effort. In addition, neural networks do not naturally provide probabilistic guarantees on their predictions. The inclusion of an external monitoring mechanism is crucial to ensure the reliability of subsequent diagnoses. This thesis attempts to solve both the problems of generalization and automatic quality assessment. The proposed solutions revolve around the development of a convolutional autoencoder, which provides a surrogate quality measure for individual segmentation masks and their generating model. In particular, we propose two different types of measures, a global score, and a pixel-wise map, and we demonstrate their use by reproducing the results of the ACDC Challenge in the absence of ground truth. Next, we integrate our autoencoder into a semi-supervised framework, capable of learning from both labeled and unlabeled data to fulfill the segmentation task.