Geometric and dosimetric evaluation of deep learning-based organs at risk auto-segmentation for rectal cancer

To evaluate the clinical applicability of deep learning based auto-segmentation of organs at risk (OARs) based on geometric and dosimetric indices. In this study, a total of 35 patients with rectal cancer who received radiotherapy in the Fudan University Shanghai Cancer Center were enrolled, and an in-house developed deep learning system was used to automatically segment OARs. Taking manual delineation as reference, the geometric evaluation indices included the dice similarity coefficient (DSC), jaccard similarity coefficient (JSC), hausdorff distance (HD), and mean distance to agreement (MDA). For each case, the treatment planning was optimized based on the auto-segmented OARs. The plan optimization and evaluation process was consistent with clinical procedure, recorded as Plan_FD. The dosimetric differences between Plan_FD and the original clinical treatment plan (recorded as Plan_Treat) were compared through dose-volume parameters and three-dimensional gamma analysis. The auto-segmented and manually delineated OARs were not only highly overlapping in three-dimensional space with an average value of DSC greater than 0.85 but also well matched in edge details with an average value of MDA less than 2.8 mm. In the dosimetric evaluation, a statistically significant difference was observed only for the bladder (p < 0.05); the dose-volume parameters of the other OARs and PTV were not statistically significant. Three-dimensional γ analysis (3 mm / 3 % standard) was performed on Plan_FD and Plan_Treat, where the percentage of points with γ ≤ 1.0 was (91.63 ± 6.27) %. Deep learning-based auto-segmentation of OARs has a high geometric similarity with manual delineation. The auto-segmented OARs have no significant impact on treatment planning and dose distribution, and have the potential to be directly applied to clinical treatment planning.