Fig. 8

Multi-class segmentations of synthetic images. Synthetic CMR images (top row, CMR), multiclass annotation from the CMRGenNet semantic segmentation branch (mid row, GAN), and MultiClassNet predictions (bottom row, UNet) are shown (a). The Dice score (DCS) value for the multi-class annotations from the semantic segmentation branch is reported. It is noted that these examples correspond to the best (first two columns), average (mid columns) and worst (last two columns) predicted cases. DCS values for the segmentation branch of the MultiClassNet on labelled images used for training and validation are provided (b). These corresponds the synthetic images generated with CMRGenNet with the corresponding multi-class masks. Labels refer to the right ventricular (RV) and left-ventricular (LV) blood pools, RV blood pool and LV blood pool, respectively, and the RV and LV myocardium (MYO), RV MYO and LV MYO, respectively. BP blood pool