Automated detection of mandibular landmarks in CT data using a dual-input approach in a two-stage design

Comprehensive Summary

This study investigates the automation of precise landmark localization in the mandible using volumetric CT imaging and convolutional neural networks. To perform this research, a pipeline was developed and tested on 287 CT datasets to localize nine different landmarks on the human mandible, including the Condyles, Coronoids, Gonions, Pogonion, Gnathion, and Menton. The pipeline was then validated on a dataset of 29 CT scans. As a result, landmarks were predicted with a mean absolute error of 1.40 +/- 1.04, with 99.6% of all landmarks being successfully predicted. Although this process does demonstrate high accuracy, robustness, and speed, it is important to note that a test data set of 29 CT scans is fairly small, reducing the presence of variation or abnormalities.

Outcomes and Implications

Current manual methods of identifying anatomical landmarks for digitized preoperative planning are time-consuming, resource-heavy, prone to inter-operator variability, and often cause a bottleneck in streamlined workflows. By automating this process, it is possible to reduce surgical costs, fast-track patient operations, and increase clinical efficiency. In addition, it would allow segmentation and landmark localization to become standardized with a reduction in intraoperative variability. Although the authors do not seem to comment on a timeline for clinical implementation, they do claim there is a “strong potential for integration into clinical workflows.” Further validation, however, is almost certainly necessary.