Assessment of deep learning-based auto-contouring on interobserver consistency in target volume and organs-at-risk delineation for breast cancer: Implications for RTQA program in a multi-institutional study

Hausdorff distance
DOI: 10.1016/j.breast.2023.103599 Publication Date: 2023-11-15T07:27:23Z
ABSTRACT
PurposeTo quantify interobserver variation (IOV) in target volume and organs-at-risk (OAR) contouring across 31 institutions breast cancer cases to explore the clinical utility of deep learning (DL)-based auto-contouring reducing potential IOV.Methods materialsIn phase 1, two were randomly selected distributed multiple for six volumes (CTVs) eight OAR. In Phase 2, auto-contour sets generated using a previously published DL Breast segmentation model made available all participants. The difference IOV submitted contours phases 1 2 was investigated quantitatively Dice similarity coefficient (DSC) Hausdorff distance (HD). qualitative analysis involved contour heat maps visualize extent location these variations required modification.ResultsOver 800 pairwise comparisons analysed each structure case. Quantitative metrics showed significant improvement mean DSC (from 0.69 0.77) HD 34.9 17.9 mm). increased agreement specifically CTV structures (5–19 %), leading fewer manual adjustments. Underlying differences causes reported questionnaire hierarchical clustering based on CTVs.ConclusionDL-based auto-contours improved OARs CTVs significantly, both qualitatively quantitatively, suggesting its role minimizing radiation therapy protocol deviation.
SUPPLEMENTAL MATERIAL
Coming soon ....
REFERENCES (32)
CITATIONS (5)