Tooth detection from low-quality camera images using deep learning technologies

Abstract

Oral health issues are global problems due to factors such as lack of finances, access to healthcare, and lifestyle. These diseases affect education and productivity, necessitating an accessible oral health care system. This thesis aimed to develop an AI-based approach for tooth detection and numbering in low quality camera images. The proposed solution involved training two models using deep learning technologies: Faster RCNN and Mask RCNN. The models were tested on a publicly available dataset in Coco format, with the Optuna optimization tool used for hyperparameter tuning. The results indicated deep learning technology is a valid tool for dental image analysis. The best performed Faster RCNN model was found at IoU threshold and prediction score threshold 0.5. It recorded precision of 92.5%, recall of 76.6% and F1 score of 83.8% in terms of tooth detection. For the same model, precision, recall, and F1 score for the label prediction were 77.3%, 36.8% and 59.8% respectively. Best performance of Mask RCNN was recorded at IoU threshold 0.5 and prediction score threshold 0.9. It achieved 91.9% of precision, 84.4% of recall and 87.8% of F1 score for detecting tooth region. For the same model, precision, recall, and F1 score for the label prediction were 80.9%, 57.4% and 67% respectively. Mask RCNN was found to be better at tooth detection and numbering than faster RCNN. However, tooth region detection by both models were better than the label prediction by both models. Generally, the models showed promising results, indicating their potential integration into the oral health care system. However, there were associated challenges such as a small dataset, incorrect labeling, and validation losses need to be implemented explicitly to achieve a more robust and accurate solution for tooth detection and numbering.