Share
Export Citation
Comparative Analysis of Backbone Architectures in Faster R-CNN for Automated Detection of Idiopathic Osteosclerosis on Panoramic Radiographs
Yudhiestra Rachman M.R.
2026 8th International Conference on Software Engineering and Computer Science Csecs 2026 Conference Proceedings
Abstract
Automatic detection of Idiopathic Osteosclerosis (IO) in panoramic dental radiographs is a challenging task due to the small size of the lesions and their visual similarity to other radiopaque lesions. In addition, most previous studies have used a limited variety of backbone architectures, so the impact of backbone selection on detection performance has not been systematically analyzed. This study presents a benchmarking analysis of nine backbone architectures integrated into the Faster R-CNN framework. The dataset consists of 297 panoramic radiographic images annotated with bounding boxes and validated by three experienced radiologists. Model performance was evaluated using sensitivity, mAP@0.5, and mAP@0.5–0.95 metrics. The experimental results show that backbone selection has a significant impact on detection performance, with the ResNet family consistently achieving strong results. Among them, ResNet-50 provided the most balanced performance, achieving the highest sensitivity of 0.5710 and a high mAP@0.5 of 0.8475, along with a competitive mAP@0.5–0.95 of 0.4949. Compared to shallower models such as ResNet-18 and ResNet-34, ResNet-50 achieves higher detection accuracy, while deeper architectures such as ResNet-101 and ResNet-152 only provide marginal improvements in certain localization metrics at the cost of substantially higher computational complexity and longer inference time. Specifically, ResNet-50 uses 41.30M parameters with an inference time of 106.11 ms, whereas deeper variants significantly increase the model size and computational cost while providing limited performance gains. These findings indicate that backbones with residual structures and balanced depth can produce more stable training convergence and more effective feature representations, thereby improving the accuracy of IO lesion detection in panoramic radiographs while maintaining a practical trade-off between detection performance and computational efficiency.