딥러닝 기반 치과 의료영상 판독에 대한 문헌 분석 (Literature Analysis of Deep Learning Based Dental Imaging Readings)

서지정보

· 발행기관 : 경희대학교 경영연구원
· 수록지 정보 : 의료경영학연구 / 14권 / 3호 / 15 ~ 28페이지
· 저자명 : 최현철, 김초명, 박상찬

초록

This study analyzes the papers, which studied to find the most adequate CNN based algorithms for segmentation, object detection in dentistry. According to our purpose, we created several keywords like “Dental+Object Detection+Neural+Network.” We searched articles in ‘PubMed’, ‘IEEE’, using created 34 keywords. We found 458 papers and excluded under a study-purpose provision. So This paper had categorized those 23 papers by 11 of segmentation of tooth structure with dental filling and FDI numbering, 12 of detecting dental caries, periodontitis, or multiple lesions. To compare the performance of models, we organized the results by DICE/IoU index and accuracy, precision, recall, etc.. Various dataset was used for analyzing. The most common dataset was dental panoramic image, then periapical, CBCT, NILT, and intra-oral image. The algorithms were used according to the purpose. For example, VGG16, 19 was used for object detection algorithms were used according to the purpose. For example, VGG16, 19 was used for object detection, U-Net, and Mask R-CNN used for segmentation by study purpose.
For segmentation of teeth, Zhimming Cui(2019), used Mask R-CNN, and the accuracy was 0.9755. Vranck(2020) used ResNet for molar detection(IoU 0.9, precision 0.94, 0.93). To label the tooth numbering according to FDI rule, Tuzoff(2019) and Chen(2019), used Faster R-CNN, VGG16, and Faster R-CNN with DNN. Tuzoff’s index was slightly better than Chen’s. Casalegno(2019) investigated the detection of dental caries by using VGG16. The result was IoU 0.727. To find periodontitis, used VGG16 also, by Prajapaty(2017). And the accuracy was 0.8846. Using the Mask R-CNN, Jader(2018) could separate instances of multiple lesions, accuracy was 0.8846.

영어초록

This study analyzes the papers, which studied to find the most adequate CNN based algorithms for segmentation, object detection in dentistry. According to our purpose, we created several keywords like “Dental+Object Detection+Neural+Network.” We searched articles in ‘PubMed’, ‘IEEE’, using created 34 keywords. We found 458 papers and excluded under a study-purpose provision. So This paper had categorized those 23 papers by 11 of segmentation of tooth structure with dental filling and FDI numbering, 12 of detecting dental caries, periodontitis, or multiple lesions. To compare the performance of models, we organized the results by DICE/IoU index and accuracy, precision, recall, etc.. Various dataset was used for analyzing. The most common dataset was dental panoramic image, then periapical, CBCT, NILT, and intra-oral image. The algorithms were used according to the purpose. For example, VGG16, 19 was used for object detection algorithms were used according to the purpose. For example, VGG16, 19 was used for object detection, U-Net, and Mask R-CNN used for segmentation by study purpose.
For segmentation of teeth, Zhimming Cui(2019), used Mask R-CNN, and the accuracy was 0.9755. Vranck(2020) used ResNet for molar detection(IoU 0.9, precision 0.94, 0.93). To label the tooth numbering according to FDI rule, Tuzoff(2019) and Chen(2019), used Faster R-CNN, VGG16, and Faster R-CNN with DNN. Tuzoff’s index was slightly better than Chen’s. Casalegno(2019) investigated the detection of dental caries by using VGG16. The result was IoU 0.727. To find periodontitis, used VGG16 also, by Prajapaty(2017). And the accuracy was 0.8846. Using the Mask R-CNN, Jader(2018) could separate instances of multiple lesions, accuracy was 0.8846.

참고자료

· 없음