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 introduced for the prostate's automatic segmentation and its coarse subdivision into transition zone (TZ) peripheral (PZ). However, when targeting a fine-grained of TZ, PZ, distal prostatic urethra (DPU) anterior fibromuscular stroma (AFS), task becomes more challenging has not yet solved at level human performance. One reason might be insufficient amount labeled data supervised training. Therefore, we propose to apply semi-supervised learning (SSL) technique named uncertainty-aware temporal self-learning (UATS) overcome expensive time-consuming manual ground truth labeling. We combine SSL techniques ensembling uncertainty-guided benefit from unlabeled images, which are often readily available. Our method significantly outperforms baseline obtained Dice coefficient (DC) up 78.9%, 87.3%, 75.3%, 50.6% DPU AFS, respectively. The results in range inter-rater performance all structures. Moreover, investigate method's robustness against noise demonstrate generalization capability varying ratios on other tasks, namely hippocampus skin lesion segmentation. UATS achieved superiority quality compared baseline, particularly minimal amounts data.