Classification of Large Periapical Lesions
Oral Defence Date:
Professors Okada, Yoon, and Petkovic
This thesis proposes a novel application of computer-aided diagnosis to a clinically significant dental problem: non-invasive differential diagnosis of periapical lesions using cone-beam CT. The proposed semi-automatic solution combines advanced segmentation and pattern classification techniques. We propose a novel robust algorithm to unify AdaBoost and LDA by introducing sample weights to the LDA formulation. Our quantitative experiments show the effectiveness of the proposed AdaBoost and Weighted LDA method by achieving higher classification accuracy than other AdaBoost combinations on multiple data sets, as well as requiring less iterations to reach a low training/test error. Furthermore, the proposed Dental CAD system achieves a 94.1\% correct classification rate, which demonstrates the validity of this approach to the clinical problem. Two independent ground-truth sets from biopsy and CBCT diagnosis are used to compare classification performance. ROC analysis reveals that our method improves accuracy for both cases and behaves more in agreement with the CBCT diagnosis than with biopsy, supporting a hypothesis presented in a recent clinical report.
Pattern classification, CAD, medical imaging