Paired Comparison of Automated and Conventional Mandibular Segmentation

Abstract

Mandibular segmentation is a critical step in three-dimensional virtual surgical planning, but conventional threshold-based segmentation can require substantial operator interaction, particularly near the dentition and temporomandibular joints. This paired study compared automated and conventional mandibular segmentation in 30 computed tomography (CT) cases from the Public Domain Database for Computational Anatomy (PDDCA). Automated segmentation was performed in 3D Slicer version 5.12.3 with TotalSegmentator version 2.14.0 using the craniofacial-structures task; the separate lower-teeth label was excluded. Conventional segmentation was performed independently by the same investigator, who had approximately six years of experience in three-dimensional medical-image processing and segmentation. The PDDCA reference was withheld during both segmentation pathways. Across the 30 paired cases, automated versus conventional segmentation yielded: mean Dice similarity coefficient was 0.821 ± 0.021 for automated segmentation and 0.828 ± 0.015 for conventional segmentation; 95th percentile Hausdorff distance (HD95) was 2.392 ± 0.241 versus 3.446 ± 0.943 mm, average symmetric surface distance (ASSD) was 0.951 ± 0.094 versus 1.152 ± 0.165 mm, and maximum symmetric surface deviation was 4.992 ± 1.051 versus 17.737 ± 8.686 mm. Mean signed volume difference was +33.06% versus +23.86%, respectively, Matched total workflow time was 23.38 ± 1.94 min for the automated pathway and 106.80 ± 13.38 min for the conventional pathway (mean paired difference = -83.42 min; 95% CI, -88.54 to -78.29 min; p < 0.001), corresponding to a 78.1% reduction.