Modified DenseNet-121 with Transfer Learning Approach for Classifying Oral Conditions

Limited access to dental care due to expensive equipment is a challenge in developing countries like the Philippines. AI-based tools offer potential for efficient and affordable oral disease diagnosis. However, previous studies using color images have shown limitations in accuracy. This study proposes a modified DenseNet-121 deep learning model for classifying common oral diseases (calculus, cavities, gingivitis) using a limited dataset (992 images). Transfer learning was employed to adapt a pre-trained DenseNet-121 architecture. Hyper parameter tuning (kernel regularization, dropout, Adam optimizer) was implemented to enhance performance on the imbalanced dataset. The modified model achieved a remarkable 95.00% accuracy, surpassing the performance reported in prior research on small datasets for oral disease classification with an increase of 13.18%. The findings indicate that deep learning models show potential for accurately classifying oral diseases with limited datasets. To verify these findings and determine its clinical usefulness, more investigation using substantial and varied datasets is necessary.