An Automated Approach for the Identification of Landmarks On Lateral Cephalograms Using Deep Neural Networks

Abstract

Cephalometric landmark identification from lateral cephalograms is a fundamental and imperative process commonly employed for orthodontic diagnosis, treatment planning, oral and maxillofacial surgeries. The emergence of completely automated landmark detection approaches in the domains of dentistry assist orthodontists and maxillofacial surgeons in accurately identifying landmarks from cephalograms. Most existing research papers utilized deep learning architectures, yielding superior outcomes in precisely localizing landmarks compared to knowledge based and machine learning approaches. A customized CephXNet framework incorporating dual convolution max pooling layers with Squeeze and Excitation attention mechanism is presented for automatic cephalometric landmark classification and prediction. The attention mechanism performs recalibration of channel characteristics dynamically by focusing on important landmark features in cephalometric X-rays and thereby models channel wise interdependency. An AdamW Golden Search Optimization based multi-level attention enhanced Stacked Feature Generator architecture named and#12310;SFGand#12311;_CephX is incorporated for 19 landmarks prediction from cephalometric X-rays. The Stacked Feature Generator model utilized multi-level attention integrated feature extractors combined with a characteristic meta-learner resulting higher Successful Detection Rates (SDRs) and minimum Mean Radial Errors (MREs) with Standard Deviations (SDs). A fused feature extraction and#12310;CephTransXand#12311;_net framework employed multiple branches to predict landmark coordinates from lateral cephalograms. The discriminative local feature extraction capability of the model is enhanced by incorporating a Sequential branch with channel and spatial attention mechanisms. The long range dependencies and global feature representation are effectively captured by Swin Transformer branch. The surrounding features of each landmark is analysed using a Landmark Discriminative Deviation Factor by applying the Neighborhood Rough Set technique. The fuse