nnU-Net Automatically Segments Mandibular Osteotomy Gaps on CBCT (Dice 0.73)

nnU-Net segmentation of osteotomy gaps after bilateral sagittal split osteotomy on CBCT had a Dice of 0.73 and minimal volumetric bias (+1.58 mm3), supporting objective healing assessment.

• On an independent hold-out set of 12 osteotomy sites, the model showed no significant volumetric difference from manual segmentation (bias +1.58 mm³).
• Reducing training data from 60 to 30 gaps maintained geometric accuracy but increased volumetric bias to +32.75 mm³, underscoring the importance of dataset size.
• The model's performance (ASSD 1.02 mm, HD95 3.52 mm) corresponds to two to three voxels of the CBCT resolution, indicating clinically acceptable accuracy.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.