Prostate zone segmentation tool

This tool automatically segments the the prostate into two zones: the central and transition zones (TZ+CZ) and the peripheral zone (PZ). It takes as input a T2-weighted image and produces a segmentation in the same input format. Three datasets were used to train this model: ProstateX (n=152), Prostate158 (n=119) and ProstateNet (the ProCAncer-I dataset; n=532). We used the T2-weighted images available in each dataset and trained a standard three-class nnU-Net model with three classes: background, peripheral zone and the combination of the transition and central zones as annotated by radiologists. Using 149 cases from the three previously noted datasets, this model achieved DSC=0.81 (95% CI=[0.59, 0.92]) for PZ and 0.87 (95% CI=[0.60, 0.96]) for CZ+TZ. In a clinical validation using ProCAncer-I data, this model achieved, for PZ and TZ+CZ, DSC=0.65 and DSC=0.77, respectively, comparable to the observed between-radiologist DSC for a subset of cases (DSC=0.61 and DSC=0.73, respectively).