Uncertainty-Aware Temporal Self-Learning (UATS): Semi-Supervised Learning for Segmentation of Prostate Zones and Beyond

Various convolutional neural network (CNN) based concepts have been\nintroduced for the prostate's automatic segmentation and its coarse subdivision\ninto transition zone (TZ) and peripheral zone (PZ). However, when targeting a\nfine-grained segmentation of TZ, PZ, distal prostatic urethra (DPU) and the\nanterior fibromuscular stroma (AFS), the task becomes more challenging and has\nnot yet been solved at the level of human performance. One reason might be the\ninsufficient amount of labeled data for supervised training. Therefore, we\npropose to apply a semi-supervised learning (SSL) technique named\nuncertainty-aware temporal self-learning (UATS) to overcome the expensive and\ntime-consuming manual ground truth labeling. We combine the SSL techniques\ntemporal ensembling and uncertainty-guided self-learning to benefit from\nunlabeled images, which are often readily available. Our method significantly\noutperforms the supervised baseline and obtained a Dice coefficient (DC) of up\nto 78.9% , 87.3%, 75.3%, 50.6% for TZ, PZ, DPU and AFS, respectively. The\nobtained results are in the range of human inter-rater performance for all\nstructures. Moreover, we investigate the method's robustness against noise and\ndemonstrate the generalization capability for varying ratios of labeled data\nand on other challenging tasks, namely the hippocampus and skin lesion\nsegmentation. UATS achieved superiority segmentation quality compared to the\nsupervised baseline, particularly for minimal amounts of labeled data.\n