2026-03-262026-03-26https://hdl.handle.net/20.500.14901/1108Dental lesion detection has a great importance for maintaining oral health and preventing the development of serious diseases. Dentists use medical imaging techniques such as panoramic, periapical and cone beam computed tomography (CIBT) for lesion detection. However, especially small lesions can be difficult to recognize by the dentist. To help dentists, deep learning models such as CNN, U-net, YOLO, DetectNet and GoogleNet have been used in the literature for automatic lesion detection. Although high accuracy rates can be achieved with these models, they require high computational costs and large data sets due to their complex structure. Transfer learning method provides lower computational cost and high performance since it can be fine-tuned using less data and a pre-trained model. Using transfer learning methods, dental panoramic radiographs will be focused on the teeth and maxillary bones and the data size will be reduced by automatically cropping the regions outside the bimaxillary structures. By splitting the image into patches, more salient features will be extracted from each patch. The resulting fragment images will be passed through Bilateral and CLAHE filters to extract features from the two most accurate network models. The models whose feature maps are combined will be passed through various feature selectors and the best performers will be determined.