Fully Automated Detection of Osteoporosis Stage on Panoramic Radiographs Using YOLOv5 Deep Learning Model and Designing a Graphical User Interface

Osteoporosis is a systemic disease that causes fracture risk and bone fragility due to decreased bone mineral density and deterioration of bone microarchitecture. Deep learning-based image analysis technologies have effectively been used as a decision support system in diagnosing disease. This study proposes a deep learning-based approach that automatically performs osteoporosis localization and stage estimation on panoramic radiographs with different contrasts. Eight hundred forty-six panoramic radiographs were collected from the hospital database and pre-processed. Two radiologists annotated the images according to the Mandibular Cortical Index, considering the cortical region extending from the distal to the antegonial area of the foramen mentale. The data were trained and validated using the YOLOv5 deep learning algorithm in the Linux-based COLAB Pro cloud environment. The Weights and Bias platform was integrated into COLAB, and the training process was monitored instantly. Using the model weights obtained, the test data that the system had not seen before were analyzed. Using the non-maximum suppression technique on the test data, the bounding boxes of the regions that could be osteoporosis were automatically drawn. Finally, a graphical user interface was developed with the PyQT5 library. Two radiologists analyzed the data, and the performance criteria were calculated. The performance criteria of the test data were obtained as follows: an average precision of 0.994, a recall of 0.993, an F1-score of 0.993, and an inference time of 14.3 ms (0.0143 s). The proposed method showed that deep learning could successfully perform automatic localization and staging of osteoporosis on panoramic radiographs without region-of-interest cropping and complex pre-processing methods.