Abstract

Info

Title:: Deep learning based software solutions for automatic segmentation of head and neck organs at risk

Author(s):: HU Xinggang1; WANG Xian1; ZHANG Yang1; ZHANG Yulei2; LI Xiaoxuan1; CHEN Meng1; 1. Cancer Center, Puer Peoples Hospital, Puer 665000, China 2. Department of Radiology, Luoyang Central Hospital, Luoyang 471000, China

Keywords:: Keywords: automatic segmentation head and neck organs at risk deep learning

Abstract:: Abstract: Objective To evaluate and analyze the accuracies of 3 software solutions based on deep learning techniques in the automatic segmentation of head and neck organs at risk (OAR). Methods The automatic segmentation accuracies of 3 software (PV-iCurve, RT-Mind, and AccuContour) were evaluated with Dice similarity coefficient (DSC), Hausdorff distance (HD), center of mass deviation (COMD), false negative rate (FNR), false positive rate (FPR), Jaccard coefficient (JC), sensitivity index (SI), and inclusive index (II) using the manual contours of head and neck OAR as the gold standard. Results The FNR, JC, SI of brain, the FPR, II of brainstem, the FPR, FNR, JC, SI of eye_L, the FPR, FNR, SI, II of mandible, the FPR, FNR, SI, II of parotid_L, and the DSC, FPR, JC, II of spinal cord manifested significant differences among the 3 software (P<0.05) but the HD, FNR, SI of brainstem, and the HD of spinal cord revealed trivial differences among the 3 software (P>0.05). Conclusion Through the comparison of multiple parameters, it is found that the accuracies of 3 software are different in OAR segmentation, which makes it difficult to make overall horizontal comparisons. Therefore, these parameters are for reference only and cannot be used as criteria for evaluating the segmentation results in clinic. Although all 3 software achieve preferable segmentation outcomes, scrutiny and manual modifications before clinical practice are still necessary.