Periapical lesion identification and pai score analysis using multiple deep learning frameworks

Objective: This study aimed to compare the performance of four deep learning (DL) architectures (DenseNet-161, EfficientNet-B0, ResNet-50, MaxViT-T) for the automated detection of periapical lesions based on Periapical Index (PAI) scores on periapical radiographs. Materials and Methods: A total of 3,541 periapical radiographs from Trakya University Faculty of Dentistry were used, yielding 6,000 tooth root images annotated by expert endodontists according to PAI scores. The dataset was divided into two main groups: PAI 1–2 (healthy) and PAI 3–4–5 (pathological). All models were fine-tuned via transfer learning and evaluated using metrics such as accuracy, sensitivity, specificity, F1-score, and area under the curve (AUC). Statistical comparisons between models were performed using the chi-square test, Bonferroni correction, and bootstrap analysis. Results: In the PAI 1–2 vs. 3–4–5 classification, DenseNet-161 achieved the highest performance (accuracy: 88%, AUC: 0.88), followed by EfficientNet-B0 (accuracy: 84%, AUC: 0.81). ResNet-50 exhibited moderate success (accuracy: 72%), whereas MaxViT-T showed poor sensitivity (29%). In the PAI 1 vs. PAI 2 classification, DenseNet-161 (accuracy: 67%, F1: 0.67) emerged as the most balanced model. Conclusion: DenseNet-161 and EfficientNet-B0 were identified as the most clinically reliable models for periapical lesion detection. This study demonstrates that AI-driven decision-support systems can enhance diagnostic accuracy and provide consistent, objective evaluation in endodontics.