Automatic Oral View Classification in the Photographic Images

Photographic images have been used to analyze oral healthcare by dentists worldwide. The oral images can be taken in different views to emphasize different anatomical structures. They can also be taken with different angles, zooms, lights, and qualities. Dentists and oral scientists have to sort these images into respective views before further analyses. This manual process was tedious, time-consuming, and error prone. It is required to establish a system that can automatically sort photographic images into one of six standard views (e.g., Buccal mucosa view, Lip views, Gingiva view, etc.) or neither. Our work here was the first comprehensive study that reviewed the possibility of using state-of-the-art algorithms to address the issue. We used 17 deep learning models that were pre-trained with millions of images in the ImageNet dataset. The results show that all models achieved a very high performance, an average of 96% in classification accuracy. Notably, the EfficientNetV2-S model performed the best, achieving an accuracy score of 98%. Interestingly, some models with relatively low numbers of parameters, such as ShuffleNet and MobileNet, demonstrated accuracy scores above 96%, with only 1.4-3.5 million trainable parameters. These models could be suitable for building an automatic oral view sorting module in resource-limited environments such as mobile applications. The successful application of these models had laid an important foundation for developing more effective and accessible diagnostic tools for oral health. We strongly believe that this study can inspire further research in the field and contribute to the continued advancement of oral healthcare.