Training biomedical image segmentation CNNs with scarce, cross-domain, prior, or noisy supervision

Artificial intelligence, and more precisely deep learning, has shown remarkable performance in the field of computer vision, which aims to automate tasks that the human visual system can do. However, the recent successes of deep learning still rely on the availability of large and high-quality annotated datasets. This is a limitation in the field of biomedical imaging, where both data and associated annotations are expensive to collect. In this thesis, four solutions are investigated to mitigate the impact of non-ideal datasets for image segmentation. We show that if an inaccurate prior segmentation is available at inference time, it can be used to guide the segmentation model and improve the segmentation accuracy of a target image. This strategy is applied to bladder segmentation in CT images. Then, we show that the problem of scarce annotations can be mitigated by using a related labeled dataset using cross-domain data augmentation. We apply cross-domain data augmentation to male pelvic organ segmentation in CBCT scans, where a small CBCT training set has been augmented with a large number of CT scans already contoured in the current clinical workflow. When a related dataset is not available, we alternatively show how to improve segmentation accuracy and better estimate the reliability of the model predictions by using multiple predictions obtained with different model initializations or multiple transformed versions of the input. We apply this methodology in the context of mineralized cartilage versus bone segmentation in high-resolution microCT scans. Finally, we consider the possibility to train a model based on automatically generated, but error-prone, labels. The use case of interest tackles the problem of cell organelles segmentation in microscopy images.