Enhancing Cephalometric Landmark Detection with a Two-Stage Cascaded CNN on Multi-resolution Multi-modal Data

Accurate identification of cephalometric landmarks in X-ray radiographs plays an integral role in orthodontic diagnosis and treatment planning. While deep-learning based approaches have demonstrated impressive capabilities to automate landmark localization, achieving robust performance and generalization remains a significant challenge. This work presents a novel two-stage cascaded convolutional neural network (CNN) architecture for automatic cephalometric landmark detection. We leverage a newly developed multi-modal, multi-resolution benchmark dataset, named “Aariz”, specifically designed for this task. The first stage employs a CNN to localize and extract the craniofacial region from cephalometric images. This pre-processing step ensures consistent feature extraction across diverse imaging conditions, thereby mitigating potential biases introduced by varying image acquisition protocols. The extracted features are then fed into a subsequent stage, where another CNN is trained to predict the locations of cephalometric landmarks using heatmap regression. Our proposed framework achieved an overall mean radial error (MRE) of 1.789 ± 6.548 on the test dataset, adequately within the clinically accepted range of 2 mm. Additionally, our framework demonstrated success detection rates (SDRs) of 78.44% and 94.44% within the 2 mm and 4 mm ranges, respectively. We have also conducted extensive ablation studies to provide deeper insights into various components of our framework. The findings highlight the promising potential of our approach, in tandem with a multi-resolution dataset, to significantly enhance orthodontic diagnosis and treatment planning through automated landmark localization.