| Abstract: |
This thesis presents an experimental study for assessing general-purpose 3D segmentation algorithms for analyzing dental periapical lesions in cone-beam computed tomography (CBCT) scans. Clinical studies have been unable to determine if a periapical granuloma can heal with non-surgical methods. Simon et al. recently proposed a technique which non-invasively classifies target lesions using CBCT. Their manual segmentation is too time consuming and unreliable for real world adoption. Presented here is a novel application of segmentation algorithms to the clinically-significant problem. This study evaluates three state-of-the-art, graph-based algorithms: normalized cut, graph cut, and random walks. I extend the 2D formulation of the algorithms to segment 3D CBCT images and experimentally evaluate the results. Furthermore, the best performing algorithm is improved using classifiability criteria. Potential thresholds are produced by extending the decision function using a likelihood ratio test. Two data-driven methods, maximum total accuracy and Bayesian cross validation criteria, quantitatively isolate optimal thresholds. These are qualitatively evaluated using 3D CBCT scans. |
