Hu Chen1,2,3,4,5,6,7, Hong Li8,9, Yijiao Zhao10,11,12,13,14,15,9, Jianjiang Zhao10,11,12,13,14,15,9, Yong Wang16,17,18,19,20,21,22. 1. Center of Digital Dentistry, Peking University School and Hospital of Stomatology, Beijing, People's Republic of China. ccen@bjmu.edu.cn. 2. Department of Prosthodontics, Peking University School and Hospital of Stomatology, Beijing, People's Republic of China. ccen@bjmu.edu.cn. 3. National Engineering Laboratory for Digital and Material Technology of Stomatology, Beijing, People's Republic of China. ccen@bjmu.edu.cn. 4. NHC Key Laboratory of Digital Technology of Stomatology, Peking University, Beijing, People's Republic of China. ccen@bjmu.edu.cn. 5. Beijing Key Laboratory of Digital Stomatology, Peking University, Beijing, People's Republic of China. ccen@bjmu.edu.cn. 6. National Clinical Research Center for Oral Diseases, No. 22 Zhongguancun Avenue South, Haidian District, Beijing, 100081, People's Republic of China. ccen@bjmu.edu.cn. 7. Shanxi Province Key Laboratory of Oral Diseases Prevention and New Materials, Taiyuan, People's Republic of China. ccen@bjmu.edu.cn. 8. Peking University Hospital of Stomatology First Clinical Division, 37A Xishiku Street, Xicheng District, Beijing, 100034, People's Republic of China. 9. Shanxi Province Key Laboratory of Oral Diseases Prevention and New Materials, Taiyuan, People's Republic of China. 10. Center of Digital Dentistry, Peking University School and Hospital of Stomatology, Beijing, People's Republic of China. 11. Department of Prosthodontics, Peking University School and Hospital of Stomatology, Beijing, People's Republic of China. 12. National Engineering Laboratory for Digital and Material Technology of Stomatology, Beijing, People's Republic of China. 13. NHC Key Laboratory of Digital Technology of Stomatology, Peking University, Beijing, People's Republic of China. 14. Beijing Key Laboratory of Digital Stomatology, Peking University, Beijing, People's Republic of China. 15. National Clinical Research Center for Oral Diseases, No. 22 Zhongguancun Avenue South, Haidian District, Beijing, 100081, People's Republic of China. 16. Center of Digital Dentistry, Peking University School and Hospital of Stomatology, Beijing, People's Republic of China. kqcadc@bjmu.edu.cn. 17. Department of Prosthodontics, Peking University School and Hospital of Stomatology, Beijing, People's Republic of China. kqcadc@bjmu.edu.cn. 18. National Engineering Laboratory for Digital and Material Technology of Stomatology, Beijing, People's Republic of China. kqcadc@bjmu.edu.cn. 19. NHC Key Laboratory of Digital Technology of Stomatology, Peking University, Beijing, People's Republic of China. kqcadc@bjmu.edu.cn. 20. Beijing Key Laboratory of Digital Stomatology, Peking University, Beijing, People's Republic of China. kqcadc@bjmu.edu.cn. 21. National Clinical Research Center for Oral Diseases, No. 22 Zhongguancun Avenue South, Haidian District, Beijing, 100081, People's Republic of China. kqcadc@bjmu.edu.cn. 22. Shanxi Province Key Laboratory of Oral Diseases Prevention and New Materials, Taiyuan, People's Republic of China. kqcadc@bjmu.edu.cn.
Abstract
OBJECTIVES: It is with a great prospect to develop an auxiliary diagnosis system for dental periapical radiographs based on deep convolutional neural networks (CNNs), and the indications and performances should be investigated. The aim of this study is to train CNNs for lesion detections on dental periapical radiographs, to evaluate performances across disease categories, severity levels, and train strategies. METHODS: Deep CNNs with region proposal techniques were constructed for disease detections on clinical dental periapical radiographs, including decay, periapical periodontitis, and periodontitis, leveled as mild, moderate, and severe. Four strategies were carried out to train corresponding networks with all disease and level categories (baseline), all disease categories (Net A), each disease category (Net B), and each level category (Net C) and validated by a fivefold cross-validation method afterward. Metrics, including intersection over union (IoU), precision, recall, and average precision (AP), were compared across diseases, severity levels, and train strategies by analysis of variance. RESULTS: Lesions were detected with precision and recall generally between 0.5 and 0.6 on each kind of disease. The influence of train strategy, disease category, and severity level were all statistically significant on performances (P < .001). Decay and periapical periodontitis lesions were detected with precision, recall, and AP values less than 0.25 for mild level, while 0.2-0.3 for moderate level and 0.5-0.6 for severe level. Net A performed similar to baseline (P > 0.05 for IoU, precision, and recall), while Net B and Net C performed slightly better than baseline under certain circumstances (P < 0.05), but Net C failed to predict mild decay. CONCLUSIONS: The deep CNNs are able to detect diseases on clinical dental periapical radiographs. This study reveals that the CNNs prefer to detect lesions with severe levels, and it is better to train the CNNs with customized strategy for each disease.
OBJECTIVES: It is with a great prospect to develop an auxiliary diagnosis system for dental periapical radiographs based on deep convolutional neural networks (CNNs), and the indications and performances should be investigated. The aim of this study is to train CNNs for lesion detections on dental periapical radiographs, to evaluate performances across disease categories, severity levels, and train strategies. METHODS: Deep CNNs with region proposal techniques were constructed for disease detections on clinical dental periapical radiographs, including decay, periapical periodontitis, and periodontitis, leveled as mild, moderate, and severe. Four strategies were carried out to train corresponding networks with all disease and level categories (baseline), all disease categories (Net A), each disease category (Net B), and each level category (Net C) and validated by a fivefold cross-validation method afterward. Metrics, including intersection over union (IoU), precision, recall, and average precision (AP), were compared across diseases, severity levels, and train strategies by analysis of variance. RESULTS: Lesions were detected with precision and recall generally between 0.5 and 0.6 on each kind of disease. The influence of train strategy, disease category, and severity level were all statistically significant on performances (P < .001). Decay and periapical periodontitis lesions were detected with precision, recall, and AP values less than 0.25 for mild level, while 0.2-0.3 for moderate level and 0.5-0.6 for severe level. Net A performed similar to baseline (P > 0.05 for IoU, precision, and recall), while Net B and Net C performed slightly better than baseline under certain circumstances (P < 0.05), but Net C failed to predict mild decay. CONCLUSIONS: The deep CNNs are able to detect diseases on clinical dental periapical radiographs. This study reveals that the CNNs prefer to detect lesions with severe levels, and it is better to train the CNNs with customized strategy for each disease.