歯科放射線

抄録を表示する抄録を非表示にする

Objective: We have developed a computer-aided diagnostic (CAD) system for osteoporosis which detects the linearly eroded endosteal margin of the lower border of the mandible on panoramic radiographs. However, the operation to determine the region of interest (ROI) has not been automated. The purpose of this study was to automate the process to determine ROIs, evaluate the validity of the position of the determined ROIs, and evaluate the diagnostic reproducibility and accuracy.
Materials and methods: Panoramic radiographs of 100 postmenopausal women were used. All of the women had completed BMD assessments of the lumbar spine. In our previous study, 14 dental practitioners had manually determined the ROIs of panoramic radiographs. Average coordinates of all images were calculated using the data of the center coordinates of ROIs. We defined them as reference coordinates. CAD was designed to determine the point on the inferior border of the mandible nearest the reference coordinate as the center of ROIs. We calculated the ratio of ROIs that were appropriately determined. If the position of an automatically determined ROI was poor, the operators determined it manually. Intra- and inter-observer reproducibility was analyzed with kappa statistics. The diagnostic efficacy (sensitivity, specificity, predictive value, and accuracy) was also analyzed.
Results: The automated CAD could detect 80.5% of all ROIs and 52.0% of all cases correctly. Intra-observer and inter-observer agreements were 0.934 and 0.918, respectively. The mean sensitivity, specificity, positive predictive value, negative predictive value, and accuracy were 80, 67, 44, 91, and 70%, respectively.
Conclusion: The improved CAD did not require manual operation in nearly half of the cases. Therefore, it was considered that the new CAD can shorten the time needed for diagnosis. A more reliable system may be possible if improvement in the accuracy of automatic ROI detection can be realized.

抄録全体を表示

PDF形式でダウンロード (467K)

J-STAGEへの登録はこちら（無料）