Quantitative Comparisons of Deep-learning-based and Atlas-based Auto- segmentation of the Intermediate Risk Clinical Target Volume for Nasopharyngeal Carcinoma.

Abstract

Manual segment target volumes were time-consuming and inter-observer variability couldn't be avoided. With the development of computer science, auto-segmentation had the potential to solve this problem. To evaluate the accuracy and stability of Atlas-based and deep-learning-based auto-segmentation of the intermediate risk clinical target volume, composed of CTV2 and CTVnd, for nasopharyngeal carcinoma quantitatively. A cascade-deep-residual neural network was constructed to automatically segment CTV2 and CTVnd by deep learning method. Meanwhile, a commercially available software was used to automatically segment the same regions by Atlas-based method. The datasets included contrast computed tomography scans from 102 patients. For each patient, the two regions were manually delineated by one experienced physician. The similarity between the two auto-segmentation methods was quantitatively evaluated by Dice similarity coefficient, the 95th Hausdorff distance, volume overlap error and relative volume difference, respectively. Statistical analyses were performed using the ranked Wilcoxon test. The average Dice similarity coefficient (±standard deviation) given by the deep-learning- based and Atlas-based auto-segmentation were 0.84 (±0.03) and 0.74 (±0.04) for CTV2, 0.79 (±0.02) and 0.68 (±0.03) for CTVnd, respectively. For the 95th Hausdorff distance, the corresponding values were 6.30±3.55 mm and 9.34±3.39 mm for CTV2, 7.09±2.27 mm and 14.33±3.98 mm for CTVnd. Besides, volume overlap error and relative volume difference could also predict the same situations. Statistical analyses showed significant difference between the two auto-segmentation methods (p<0.01). Compared with the Atlas-based segmentation approach, the deep-learning-based segmentation method performed better both in accuracy and stability for meaningful anatomical areas other than organs at risk.