Automatic segmentation of the prostate on CT images using deep neural networks (DNN).

Document Type


Publication Date


Publication Title

International journal of radiation oncology, biology, physics


PURPOSE: Recent advances in deep neural networks (DNN) have unlocked opportunities toward its application for automatic image segmentation. We have evaluated a DNN-based algorithm for automatic segmentation of the prostate gland on a large cohort of patient images.

MATERIALS AND METHODS: Planning-CT (pCT) datasets for 1114 prostate cancer patients were retrospectively selected and divided into 2 groups. Group A contained 1104 datasets, with 1 physician-generated prostate gland contour for each dataset. Among these image sets, 771/193/140 were used for training, validation and testing respectively. Group B contained 10 datasets; each including prostate contours delineated by 5 independent physicians, and a consensus contour generated using the STAPLE method in CERR. All images were resampled to spatial resolution of 1x1x1.5 mm. A region (128x128x64 voxels) containing the prostate was selected to train a DNN. The best-performing model on the validation datasets was used to segment the prostate on all testing images. Results were compared between DNN and physician-generated contours using the Dice coefficient (DSC), Hausdorff distances, regional contour distances, and center-of-mass (COM) distances.

RESULTS: Mean DSC between DNN-based prostate segmentation and physician-generated contours for test data in group A, group B, and group B-consensus was 0.85±0.06 [range=0.65, 0.93], 0.85±0.04 [range=0.80, 0.91], and 0.88±0.03 [range=0.82, 0.92] respectively. The Hausdorff distance was 7.0±3.5 mm, 7.3±2.0 mm, and 6.3±2.0 mm for group A, group B, and group B-consensus respectively. The mean COM distances for all 3 dataset groups were within 5 mm.

CONCLUSIONS: A deep-neural-network-based algorithm was used to automatically segment the prostate for a large cohort of prostate cancer patients. DNN-based prostate segmentations were compared to the consensus contour for a smaller group of patients; the agreement between DNN segmentations and consensus contour was similar to the agreement reported in a previous study. Clinical use of DNN is promising, but further investigation is warranted.

PubMed ID



ePub ahead of print