Investigation of the best effective fold of data augmentation for training deep learning models for recognition of contiguity between mandibular third molar and inferior alveolar canal on panoramic radiographs
Issued Date
2023-01-01
Resource Type
ISSN
14326981
eISSN
14363771
Scopus ID
2-s2.0-85152463291
Journal Title
Clinical Oral Investigations
Rights Holder(s)
SCOPUS
Bibliographic Citation
Clinical Oral Investigations (2023)
Suggested Citation
Papasratorn D., Pornprasertsuk-Damrongsri S., Yuma S., Weerawanich W. Investigation of the best effective fold of data augmentation for training deep learning models for recognition of contiguity between mandibular third molar and inferior alveolar canal on panoramic radiographs. Clinical Oral Investigations (2023). doi:10.1007/s00784-023-04992-6 Retrieved from: https://repository.li.mahidol.ac.th/handle/20.500.14594/81819
Title
Investigation of the best effective fold of data augmentation for training deep learning models for recognition of contiguity between mandibular third molar and inferior alveolar canal on panoramic radiographs
Author's Affiliation
Other Contributor(s)
Abstract
Objectives: This study aimed to train deep learning models for recognition of contiguity between the mandibular third molar (M3M) and inferior alveolar canal using panoramic radiographs and to investigate the best effective fold of data augmentation. Materials and methods: The total of 1800 M3M cropped images were classified evenly into contact and no-contact. The contact group was confirmed with CBCT images. The models were trained from three pretrained models: AlexNet, VGG-16, and GoogLeNet. Each pretrained model was trained with the original cropped panoramic radiographs. Then the training images were increased fivefold, tenfold, 15-fold, and 20-fold using data augmentation to train additional models. The area under the receiver operating characteristic curve (AUC) of the 15 models were evaluated. Results: All models recognized contiguity with AUC from 0.951 to 0.996. Ten-fold augmentation showed the highest AUC in all pretrained models; however, no significant difference with other folds were found. VGG-16 showed the best performance among pretrained models trained at the same fold of augmentation. Data augmentation provided statistically significant improvement in performance of AlexNet and GoogLeNet models, while VGG-16 remained unchanged. Conclusions: Based on our images, all models performed efficiently with high AUC, particularly VGG-16. Ten-fold augmentation showed the highest AUC by all pretrained models. VGG-16 showed promising potential when training with only original images. Clinical relevance: Ten-fold augmentation may help improve deep learning models’ performances. The variety of original data and the accuracy of labels are essential to train a high-performance model.