AUTOMATIC CLASSIFICATION SYSTEM FOR PERIAPICAL LESIONS WITH THE TOOTHNET CNN ARCHITECTURE INTEGRATING THE PROPOSED HYBRID ACTIVATION FUNCTION ON CBCT SCANS

Imaging techniques are widely used in dentistry to understand the 3D structure of teeth and detect diseases, but their interpretation is time-consuming and prone to error. To address this, decision support systems are increasingly utilized. This study proposes a CNN-based classification model using the UFPE dataset, which includes Cone Beam Computed Tomography (CBCT) scans. In the first scenario, both real and enhanced images were input into a CNN, yielding 68.92% accuracy for enhanced images. Due to a result, enhanced images were used in all other scenarios. In the second scenario, a newly designed CNN architecture called ToothNet, incorporating a custom activation function, was tested. It achieved 69.92% accuracy, 61.45% recall, 62.67% precision, and 68.68% F1-score, showing a 1.45% increase in accuracy. To evaluate generalizability, three more classification scenarios were examined using the same dataset. ToothNet achieved 80.14% accuracy in the “healthy vs. large lesion” and 68.73% in the “healthy vs. small lesion” classification. These results indicate that the proposed architecture not only improves accuracy but is also generalizable across different lesion sizes.