Back to all papers

Deep learning-based CT segmentation of mandibular bone invasion in oral squamous cell carcinoma.

November 21, 2026pubmed logopapers

Authors

Steybe D,Fiederling E,Poxleitner P,Otto S,Obermeier KT,Sabel B,Rau S,Reisert M,Russe MF

Affiliations (5)

  • Department of Oral and Maxillofacial Surgery and Facial Plastic Surgery, LMU Medizin, LMU University Hospital, Ludwig-Maximilians-Universität München, Lindwurmstr. 2a, Munich, 80337, Germany. [email protected].
  • Department of Oral and Maxillofacial Surgery and Facial Plastic Surgery, LMU Medizin, LMU University Hospital, Ludwig-Maximilians-Universität München, Lindwurmstr. 2a, Munich, 80337, Germany.
  • Department of Radiology, LMU Medizin, LMU University Hospital, Ludwig- Maximilians-Universität München, Munich, Germany.
  • Department of Diagnostic and Interventional Radiology, Faculty of Medicine, Medical Center - University of Freiburg, University of Freiburg, Freiburg, Germany.
  • Department of Medical Physics, Faculty of Medicine, Medical Center - University of Freiburg, University of Freiburg, Freiburg, Germany.

Abstract

Accurate identification of mandibular invasion by oral squamous cell carcinoma (OSCC) is important for virtual surgical planning, but manual delineation on CT is challenging and time-consuming. This retrospective proof-of-concept study evaluated automated segmentation of the mandible and radiologically apparent bone invasion and explored centerline-based assessment of longitudinal tumor extent for computer-aided design/computer-aided manufacturing (CAD/CAM) planning. Fifty preoperative CT examinations from 50 patients with histopathologically confirmed mandibular invasion were included. Thirty examinations were used for model development and 20 for independent testing. A three-dimensional MedNeXt architecture implemented within nnU-Net was trained applying expert-reviewed CT reference segmentations. Performance was assessed using Dice similarity coefficient (DSC), precision, recall, average symmetric surface distance and 95th-percentile Hausdorff distance. Spatial localization, longitudinal invasion-span differences and signed centerline-projected endpoint errors were assessed in the holdout cohort. Mean mandibular DSC was 0.965 (95% confidence interval [CI], 0.961-0.969; median 0.967). Mean invasion DSC was 0.553 (95% CI, 0.425-0.660; median 0.623), with mean precision of 0.742 and recall of 0.483. Predicted invasion overlapped the CT reference in 18/20 examinations (90%); 16/20 (80%) achieved DSC ≥ 0.20. Two cases were completely missed. Among the 18 evaluable cases, the median absolute invasion-span difference was 7.01 mm (interquartile range, 5.08-13.00 mm) and the median symmetric endpoint distance was 5.88 mm (3.19-9.16 mm). The present proof-of-concept study demonstrates the technical feasibility of automated CT segmentation of the mandible and radiologically apparent mandibular invasion by OSCC. The approach may support transmission of radiological tumor extent into virtual resection planning. Larger multicenter studies and prospective clinical validation are needed to establish its impact on planning accuracy and workflow efficiency.

Topics

Deep LearningTomography, X-Ray ComputedCarcinoma, Squamous CellMouth NeoplasmsMandibleMandibular NeoplasmsJournal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAISlice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.