An End-to-End Deep Learning Pipeline for Automated Mandible Virtual Surgical Planning Using Real-World Clinical Data
Issued Date
2026-01-01
Resource Type
eISSN
21693536
Scopus ID
2-s2.0-105041821780
Journal Title
IEEE Access
Rights Holder(s)
SCOPUS
Bibliographic Citation
IEEE Access (2026)
Suggested Citation
Kamboonsri N., Sriwaranon T., Tantisereepatana N., Puncreobutr C., Lohwongwatana B., Olson G.B., Tel A., Robiony M., Achakulvisut T., Vateekul P. An End-to-End Deep Learning Pipeline for Automated Mandible Virtual Surgical Planning Using Real-World Clinical Data. IEEE Access (2026). doi:10.1109/ACCESS.2026.3702327 Retrieved from: https://repository.li.mahidol.ac.th/handle/123456789/117444
Title
An End-to-End Deep Learning Pipeline for Automated Mandible Virtual Surgical Planning Using Real-World Clinical Data
Corresponding Author(s)
Other Contributor(s)
Abstract
Virtual surgical planning (VSP) is essential step in mandible reconstruction, involving multiple tasks, such as structure segmentation, cutting-plane definition, and implant planning. While many studies have shown promising results on subtasks using specialized ML models, these models are typically developed in isolation. This fragmentation prevents their seamless integration into a unified, automated clinical pipeline. We present a fully automated VSP assistant integrated into 3D Slicer combining segmentation, reconstruction, and planning modules into a single clinical framework. The proposed pipeline was validated on real clinical datasets, including unhealthy mandible dataset containing mandible abnormalities, healthy mandible dataset with normal mandible, and fibula dataset of normal fibula. Quantitative evaluation using the dice similarity coefficient (DSC) yielded scores of 0.911 for mandible segmentation, 0.854 for defect segmentation, 0.908 for fibula segmentation, and 0.816 for mandible reconstruction. Planning module reduced average operation time by 57% while maintaining accuracy comparable to manual planning. End-to-end evaluation on real clinical cases demonstrated a dramatic reduction in the total time engineer spent completing tasks from average of 5 hours 21 minutes using the manual workflow to 29 minutes 38 seconds. Our results demonstrate that automated VSP offers a practical, clinically precise solution that reduces the time-intensive nature of VSP, bridging the gap between isolated ML research and real-world surgical practice.