Detection of Cavities in Radiographs Using Machine Learning: Approaches for Limited Data

Detecting cavities is crucial for diagnosing and preventing their progression and potential damage to teeth. However, early-stage cavities are small, symptomless, and can form in hard-to-reach areas, making them difficult to identify. Traditional diagnostic methods like visual inspection and dental X-rays have limitations, as certain characteristics are not visible to the naked eye. Intelligent systems for cavity detection can improve complex diagnoses in such cases. However, training algorithms for cavity detection is challenging due to limited access to labeled databases for security reasons. This study proposes training object detection algorithms with a reduced amount of data, including smaller subsets. Two established model architectures, YOLOv5 and Faster R-CNN, were used. They were trained with predefined ranges of data quantities (20%, 40%, 60%, 80%, and 100% of the total training data). A database containing 936 radiographic images with annotations for cavities and periodontal pockets was used, and only images with cavities were considered, resulting in 256 images. Results show that the YOLOv5 model achieved 85.71% precision and 60.00% sensitivity using 40% of the available data, while Faster R-CNN did not perform as well. This study demonstrates that even with limited data, it is possible to train a satisfactory object detection model for cavity detection.