Figure 6.

Representative results on 10 test datasets by using a model based on U-Net-5 trained on 7 original MRIs (U-Net-5-trained-7) and by using a model based on U-Net-5 trained on 14 original MRIs (U-Net-5-trained-14). Original slice, prostate (ground-truth and two predictions), tumor (ground-truth and two predictions), background (ground-truth and two predictions). Each row corresponds to a patient and the columns to the original image, ground truth of the wall, predicted wall by U-Net-5-trained-7, predicted wall by U-Net-5-trained-14, ground-truth of the tumor, predicted tumor by U-Net-5-trained-7, predicted tumor by U-Net-5-trained-14, ground-truth of the background, predicted background by U-Net-5-trained-7 and predicted background by U-Net-5-trained-14, respectively. For each patient the slice corresponding to the best tumor segmentation by U-Net-5-trained-7 was plotted. (The first column corresponds to the original slice, the second column to the ground truth of the wall mask, tumor mask and background). For each patient the slice corresponding to the best tumor segmentation obtained by the UNet-5 model trained on 7 original MRIs was plotted. Values are the dice coefficients for the plotted slice).