ARTIFICIAL INTELLIGENCE MODELS FOR THE RADIOGRAPHIC DETECTION OF ROOT CANAL TREATMENT IN ENDODONTICS: A COMPARISON OF YOLOV8 AND CNN-BASED DEEP LEARNING MODELS USING CLAHE-BASED IMAGE ENHANCEMENT AND EXPLAINABLE AI APPROACHES
X-International Antalya Scientific Research and Innovative Studies Congress, Antalya, Turkey, 24 - 25 May 2026, pp.113-121, (Full Text)
- Publication Type: Conference Paper / Full Text
- City: Antalya
- Country: Turkey
- Page Numbers: pp.113-121
- Open Archive Collection: AVESIS Open Access Collection
- Kütahya Health Sciences University Affiliated: Yes
Abstract
ABSTRACT
The accurate detection of root canal treatment on panoramic dental radiographs plays a significant role in endodontic assessment and clinical decision-making processes. Particularly in clinics with high patient volumes, the manual examination of radiographs can lead to time wastage, and overlooked findings may result in incorrect clinical decisions. In recent years, artificial intelligence and deep learning-based methods have shown significant advancements in the field of dental imaging. In this study, the performance of YOLOv8 and CNN-based deep learning models was compared for the automatic detection of root canal treatment. The study utilised the open-access Dental X-Ray Panoramic Dataset. The dataset comprises a total of 27,864 panoramic dental images and 30 distinct classes. For the purposes of this study, the dataset was reorganised and converted into a two-class structure: “root canal treatment present” and “root canal treatment absent”. The dataset was analysed by dividing it into training, validation and test subsets. In the CNN-based approach, the MobileNetV2 model was used with a transfer learning method; as a second approach, the YOLOv8n-cls classification model was trained. Furthermore, the CLAHE (Contrast Limited Adaptive Histogram Equalisation) method was applied during the image pre-processing stage, and the decision-making processes of the models were visualised using a Grad-CAM-based explainable AI approach. Model performances were evaluated using the following metrics: Accuracy, Precision, Recall, Specificity, F1-score, Balanced Accuracy, MCC and ROC-AUC. According to the results obtained, the YOLOv8n-cls model without CLAHE achieved the highest performance ( ). The Accuracy value was calculated as 0.944, the F1-score as 0.928, and the ROC-AUC as 0.977. The MobileNetV2 model without CLAHE, on the other hand, achieved an accuracy of 0.912, an F1-score of 0.881, and an ROC-AUC of 0.963. It was observed that the application of CLAHE did not improve performance in either model. This may be due to the fact that root canal fillings already exhibit distinct radiopaque structures in panoramic radiographs. GradCAM analyses demonstrated that the CNN model focused particularly on root canal filling regions when making decisions. Consequently, it was concluded that the YOLOv8-based approach offers higher performance in the automatic detection of root canal treatment and is promising for clinical screening applications.
Keywords: Endodontics, Deep Learning, YOLOv8, CNN, MobileNetV2, CLAHE, Explainable Artificial Intelligence, Dental Radiography.