Quantitative results of synthetic data.

PSNR values presented as boxplots, calculated between real image data and synthetic image data generated from corresponding silver truth and segmentation masks. Those silver truth segmentations are generated with automated approaches and validated to be reliable for training purposes. Whiskers range from the 5th to the 95th quantile, median values are indicated as orange line while mean values are depicted as green triangle and boxes represent the interquartile range. The involved datasets are 3D Caenorhabditis elegans (CE) [ 19 ], 2D Mouse Stem Cells (GOWT1) [ 19 ], 2D HeLa Cells (HeLa) [ 19 ], 3D Nuclei and Membranes of Danio rerio (DRNuc,DRMem) [ 21 ], 2D+t mitotic progression in Mouse Stem Cells (HeLa+t) [ 24 ] and 2D overlapping cervical cancer cells (Cerv) [ 39 ]. Using corresponding sketches and the optimized settings of t start = 400 and σ = 1, the backward process was used to replicate corresponding image samples and PSNR values were calculated to assess similarity between synthetic and real versions. The Danio rerio multi-channel data and the temporal HeLa data present special cases, which demonstrate limitations of the proposed approach. (PDF)