IntroductionThree-dimensional mandibular models are useful for planning maxillofacial surgery and orthodontic treatment. 1,2 In studies of growth, mandibular models are important for assessing morphological changes over time. 3,4 Such models are typically obtained from conventional computed tomography (CT), using high radiation dose to capture fine detail of the bony structure. Cone beam computed tomography (CBCT) shows promise for oral and craniofacial imaging applications due to lower radiation dose, lower cost and shorter acquisition time compared to CT. However, CBCT images have lower contrast and higher levels of noise than conventional CT, making mandible segmentation a challenging task. 5 Segmenting a 3D mandible is typically done 'interactively' in computer software on a case by case basis. Threshold-based algorithms or morphological operations are commonly used first for the separation of bony structures from soft tissues.6-8 Then manual work is needed to separate the mandible from the cranial base and the maxilla because the algorithms cannot distinguish between different facial bones with similar intensity values. We refer to this combination of computerized operations and manual editing as 'interactive' segmentation. Specific issues in mandible segmentation include intercuspation occlusion and low contrast of condyles relative to surrounding structures. Intercuspation leads to connection between upper and lower teeth while low contrast leads to difficult to define boundaries on the condyles. These require slice-by-slice based manual editing, which is tedious, time-consuming and operator-dependent as it produces slightly different outlines after continuous interventions.
6A robust automated mandible segmentation approach is thus desirable. In other applications, simple automated methods based on voxel intensity or edge intensity can be very effective.9,10 However, the mandible is not the only bone structure in CBCT All rights reserved. No reuse allowed without permission.(which was not peer-reviewed) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity.The copyright holder for this preprint . http://dx.doi.org/10.1101/397166 doi: bioRxiv preprint first posted online Aug. 21, 2018; images of the head, and the intensity of bone varies considerably. More sophisticated methods which incorporate prior information about the expected shapes and position of objects to be segmented are required. To date only four publications, using statistical shape models, 11,12 multi-atlas label registration 13 or machine learning, 14 have been proposed to automate the mandibular segmentation from CBCT images. These approaches are either computationally expensive or require collection of large amounts of manually segmented mandibles as training data, which may be impractical in clinical situations.The watershed method is a classic, computationally simple technique for object segmentation in images. The original grayscale image can be regarded as a topographic relief, with brightne...