Comparison of automated segmentation techniques for magnetic resonance images of the prostate

Abstract Background: The contouring of anatomical regions is a crucial step in the medical workflow and is both time-consuming and prone to intra- and inter-observer variability. This study compares different strategies for automatic segmentation of the prostate in T2-weighted MRIs.Methods: This study included 100 patients diagnosed with prostate adenocarcinoma who had undergone multi-parametric MRI and prostatectomy. From the T2-weighted MR images, ground truth segmentation masks were established by consensus from two expert radiologists. The prostate was then automatically contoured with six different methods: (1) an ad-hoc multi atlas-based algorithm, (2) a commercial package named SyngoVia, and four deep learning models: (3) a U-net deep learning architecture trained from scratch, (4) a U-net with a pre-trained encoder, (5) a GAN extension of the network in point 4, and (6) a segmentation-adapted modification of Google Brain’s EfficientDet architecture. The resulting segmentations were compared and scored against the ground truth masks with one 50/50 and one 70/30 train/test data split.Results: The best performing method was the adapted EfficientDet (model 6), achieving a mean Dice coefficient of 0.914, a mean absolute volume difference of 5.9%, a mean surface distance (MSD) of 1.93 pixels, and a mean 95th percentile Hausdorff distance of 3.77 pixels. The deep learning-based models were more reliable in terms of worst-case performance (0.854 minimum Dice and 4.02 maximum MSD), and no significant relationship was found between segmentation performance and clinical variables.Conclusion: Deep learning-based segmentation techniques can consistently achieve Dice coefficients of 0.9 or above with as few as 50 training patients, regardless of architectural archetype. Our atlas-based method and commercial software performed significantly worse (0.855-0.887 Dice).