Automatic segmentation of the prostate gland on planning ct images using deep neural networks (DNN).
Recommended Citation
Liu C, Gardner S, Wen N, Siddiqui F, Movsas B, and Chetty I. Automatic segmentation of the prostate gland on planning ct images using deep neural networks (DNN). Med Phys 2018; 45(6):e464.
Document Type
Article
Publication Date
2018
Publication Title
Med Phys
Abstract
Purpose: Recent advances in deep neural networks (DNN) have opened the doors toward application of DNN 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. Methods: Planning-CT (pCT) datasets for 1114 prostate cancer patients were retrospectively selected and divided into 2 groups. Group A contained 1104 datasets, with prostate gland contours delineated by physician on each dataset. Among these image sets, 964 were used for training/validation and 140 were used for testing. Group B contained 10 datasets; each including prostate contours delineated by 5 independent physicians, and consensus contour generated with the STAPLE method in CERR to combine the 5 physician contours. All images were resampled to spatial resolution of 1 mm × 1 mm × 1.5 mm. A region (128 × 128 × 64) was selected in the vicinity of the prostate to train a DNN. The top performing DNN was chosen based on validation results, and used to segment the prostate on all testing images. Results were compared between DNN and physician-generated contours using Dice coefficient (DSC). Results: Mean DSC between automatic DNN-based prostate segmentation and physician-delineated contours for test data in group A was 0.85 + 0.06 [range = 0.65, 0.93]. Mean DSC between automatic DNN-based prostate segmentation and the 5 physician contours in group B was 0.85 + 0.04 [range = 0.80, 0.91]. Mean DSC between automatic DNN-based prostate segmentation and the consensus contour in group B was 0.88 + 0.03 [range = 0.82, 0.92]. Conclusion: A deep-neural-network-based algorithm was used to automatically segment the prostate for a large cohort of prostate cancer patients. Furthermore, the DNN prostate segmentations were compared to consensus contour for a smaller group of patients; the agreement between DDN segmentations and consensus contour was similar to the agreement reported in a previous study between clinician contours and consensus. Clinical use of DNN is promising, but further investigation is warranted.
Volume
45
Issue
6
First Page
e464
