Oral Tissue Detection in Photographic Images Using Deep Learning Technology

Oral tissue detection is an important task in many image-based computer aided analyses of oral healthcare. By detecting the shape of oral tissue in the photographic image, one can eliminate other non-essential parts, which can help optimize the further processes in the pipeline, for example, oral lesion detection, cancer classification, oral tissue segmentation, and dental health analyses. Manual labeling of data by humans can be inefficient, time-consuming, and error-prone, due to subjectivity, intra- and/or inter- observer variability between experts. Therefore, in this paper, we aimed to comprehensively evaluate deep learning-based models for detecting the oral tissue in various perspectives of photographic images. We studied four different state-of-the-art object detection models including Faster R-CNN, RetinaNet, SSD, and YOLOv5. Our models provided good results in the oral tissue detection application, with the Intersection over Union (IoU) over 90 % and F1-Score over 97%. We found that the Faster R-CNN with ResNet50 backbone and RetinaNet with ResNet50 backbone were the most accurate models in detecting objects. Alternatively, Faster R-CNN with MobileNetV3 backbone and SSD with VGG16 backbone were the best models in terms of processing speed, making them well-suited for handling large datasets. We hope that our findings here contribute to the improvement of the image-based oral healthcare analysis system in the future.