Diagnosis of oral diseases in images based on machine learning

Abstract

Automated detection of oral diseases is a crucial challenge in dentistry. This study analyzes the impact of advanced data augmentation techniques and proper hyperparameter tuning on the performance of the YOLOv8 model for the accurate identification of oral diseases. Following the CRISP-DM methodology, 3,814 labeled images were processed using transformations such as rotations, brightness and contrast adjustments, and Gaussian noise to enhance variability and robustness during training. Additionally, key hyperparameters such as learning rate and batch size were optimized. The results showed that combining data augmentation with optimal hyperparameter adjustments significantly improved the model’s performance, achieving an accuracy of 89.10%, sensitivity of 86.07%, F1-Score of 87.56%, and an mAP@50 of 90.70%. Similarly, an mAP@50-95 of 56.40% was obtained, surpassing the standard configurations used in previous studies. The stable convergence of the training curves validates the effectiveness of the applied strategies and the model's ability to generalize properly. These findings highlight the importance of implementing advanced preprocessing and optimization techniques in automated oral disease diagnosis, enhancing precision and enabling early detection in clinical practice.