Title: Voxel wise segmentation of CBCT images to identify the inferior alveolar canal
Abstract:
One of the main issues in a mandibular surgery as an oral surgeon is understanding the anatomy of a patient. In particular, the inferior alveolar canal (IAC) is the single most important feature considered during these types of surgeries. This is accomplished through direct palpation as well as x-rays. However, x-rays are limited in information because of the 3D nature. By developing 3D CBCT segmentation models for the IAC, we can allow the surgeon to have access to more precise information about the patient’s anatomy. The rationale of this experiment is to find ways to better aid them in understanding the anatomy of a particular patient and be able to visualize it before a surgery. Previous papers investigating this topic were limited in their dataset, in terms of size and accuracy. This experiment works to build off a newly constructed dataset by Cipriano, et al. Limited testing of this dataset with state of the art models has been conducted. This experiment further serves to help find more accurate models. This study identified a generic U-net architecture model utilizing intersection of Union and Dice Coefficient as its loss function in the analysis of the dataset provided by Cipriano et al. as having a final evaluation of 93.3334%. This model showed improvement using intersection of Union (IoU) loss and Dice Coefficient loss, and the results are better than previous models used on this dataset.
