Development of computer aided diagnosis software for treatment evaluation of periapical lesions

Abstract

Intraoral radiographs have been commonly taken to diagnose periapical lesions and to evaluate the effectiveness of subsequent endodontic treatments. However, visual interpretation is often difficult due to various imaging factors influenced by different projection geometry and dynamic ranges of digital images. Therefore, we sought to develop an imaging based quantitative model to evaluate periapical lesions and their treatments. In the image representing before treatment, four landmarks were interactively marked and the lesion area was selected manually as region of interest (ROI). With the help of the landmarks, image registration was performed by affine transformation that transferred the ROI to the image representing after treatment. To normalize the image, the dentin area and the background of periapical image were used as references. Trabecular bone images were processed by erosion, tophat operation, and recombined in multiple scales. Intensity and morphology features were calculated based on the segmented images within the ROI. Texture features were calculated from the grayscale image within the ROI. Significant features were found in effectively treated cases (n=46) by Wilcoxon pairwise signed rank test of the paired images representing before and after treatments. Relative differences were calculated from features representing before and after treatments. Several evaluation models were derived by logistic regression analysis and support vector machine (SVM), and tested by 10-fold cross validation. Intensity and texture features were significantly increased in the effectively treated cases. A trabecular bone area feature was increased. Bone marrow area and volume features were decreased. Skeleton features of trabecular bone and bone marrow were increased. The logistic regression model provided the sensitivity of 80.0% and the specificity of 84.8%, whereas SVM resulted in 86.7% and 93.5%, respectively. Our quantitative models may be helpful to evalu...