Evaluation of clinical and radiographic findings with artificial intelligence algorithms for detecting periapical lesion

Aim: The detection of teeth and diseases using artificial intelligence in two-dimensional dental radiographs is one of the popular research areas. The identification and interpretation of periapical lesions is a time-consuming and error-prone task for dentists. This study aimed to detect the presence of periapical lesions in two different imaging methods, panoramic radiography and periapical radiography, and to evaluate the performance of various hybrid models in lesion detection. Materials and methods: In this prospective study, 800 panoramic and 800 periapical radiographic images supported by clinical findings of the patients were included. Six different ROI areas were cropped from panoramic radiographs, and four different ROI areas from periapical radiographs. Deep learning models VGG-16 and VGG-19 were used for feature extraction, and hybrid models that combined machine learning algorithms such as Logistic Regression (LR) and Support Vector Machines (SVM) were employed (VGG-16+SVM, VGG-16+LR, VGG-19+SVM, VGG-19+LR). Five different random splits (% training data - test data: 90%-10%, 85%-15%, 80%-20%, 75%-25%, 70%-30%) and three different cross-validation methods (3-fold, 5-fold, 10-fold cross-validation) were used for dataset separation. Results: In this thesis, the highest results for periapical lesion detection in panoramic radiographs were obtained from ROI 4 (200X180) using the VGG-16 model with LR machine learning, yielding a 10-fold cross-validation combination; accuracy: 0.87, precision: 0.87, sensitivity: 0.87, and F1 score: 0.86. For periapical lesion detection in periapical radiographs, the highest results were obtained from ROI 2 (300X500) using the VGG-16 model with SVM machine learning, achieved through the random split-10 combination; accuracy: 0.91, precision: 0.92, sensitivity: 0.91, and F1 score: 0.91. Conclusion: Artificial intelligence hybrid models achieved successful results in detecting periapical lesions in both panoramic and periapical radiographs. It is considered that these AI hybrid models could serve as a diagnostic system that assists dentists in diagnosis and provides time savings. Keywords: Periapical lesion, Panoramic radiograph, Periapical radiograph, Artificial intelligence, Hybrid model