Automatic Detection of Craniomaxillofacial Anatomical Landmarks on CBCT Images Using 3D Mask R-CNN.

Craniomaxillofacial (CMF) landmark localization is an important step for characterizing jaw deformities and designing surgical plans. However, due to the complexity of facial structure and the deformities of CMF patients, it is still difficult to accurately localize a large scale of landmarks simultaneously. In this work, we propose a three-stage coarse-to-fine deep learning method for digitizing 105 anatomical craniomaxillofacial landmarks on cone-beam computed tomography (CBCT) images. The first stage outputs a coarse location of each landmark from a low-resolution image, which is gradually refined in the next two stages using the corresponding higher resolution images. Our method is implemented using Mask R-CNN, by also incorporating a new loss function that learns the geometrical relationships between the landmarks in the form of a root/leaf structure. We evaluate our approach on 49 CBCT scans of patients and achieve an average detection error of 1.75 ± 0.91 mm. Experimental results show that our approach overperforms the related methods in the term of accuracy.