Faster R-CNN Based Cephalometric Landmarks Detection

Abstract

The objective of this study is to identify cephalometric landmarks on 2D cephalograms (X-rays) using a two-stage Artificial Intelligence (AI) based object detection method. The proposed work implements a Faster Region-based Convolutional Neural Network (Faster R-CNN), a deep-neural network, which consists of a 50 layered Residual Network (ResNet50) with Feature Pyramid Network (FPN) as a backbone network. The algorithm is trained and tested on a dataset presented in the IEEE International Symposium on Biomedical Imaging Challenge (ISBI-2015). The detection was based on the algorithm’s performance, in terms of mean error and the success rate under the clinically accepted accuracy range of 2 mm. The hypothesis behind this work was that Faster R-CNN will have a difficulty in detecting the landmarks due to either fuzzy features and, or low-resolution representations, but with help of FPN, the performance might be better. Results show that the model achieves approximately 90% and 0.9 mm in terms success rate and mean error respectively. In terms of future work, there is still a need to improve Faster R-CNN performance by increasing or modifying the dataset. Furthermore, the use of a more powerful computational platform would lead to faster training time, which would give room to the implementation of optimization algorithms of the hyper parameters by using evolutionary computation methods.