IEEE Access (Jan 2023)

CephXNet: A Deep Convolutional Squeeze-and-Excitation Model for Landmark Prediction on Lateral Cephalograms

  • R. Neeraja,
  • L. Jani Anbarasi

DOI
https://doi.org/10.1109/ACCESS.2023.3307636
Journal volume & issue
Vol. 11
pp. 90780 – 90800

Abstract

Read online

Cephalometric landmark identification is a crucial and significant procedure that is generally used for orthodontic treatment planning and diagnosis. Computer-aided fully automated solutions can assist orthodontists and orthognathic surgeons’ to precisely identify the landmarks from cephalograms more efficiently. Most of the existing research studies deployed Convolutional Neural Network models, transfer learning methods, and pre-trained architectures to predict the XY coordinates of landmarks from cephalometric radiographs. Deep learning architectures have achieved good results in accurately predicting landmarks compared to machine learning methodologies. In this paper, a custom CNN model integrated with Squeeze-and-Excitation (SEB) attention block named CephXNet architecture is proposed to automatically classify and predicts the XY coordinates of 19 landmarks from lateral cephalograms. The SEB block, which can adaptively refine information from various feature channels of X-ray images models the interdependence between the channels to enhance the discriminative features and suppress noise. The Squeeze-and-Excitation block incorporated with multiple Convolution and Max pooling layers enhances independent channel feature learning and thereby improves the representational power of the proposed CephXNet architecture. The proposed framework obtained an accuracy of 97.72% for 19 landmark classifications and has achieved 88.06% and 78.72% of Successful Detection Rate (SDR) in clinically accepted 2mm precision range for test1 and test2 datasets provided in the 2015 International Symposium on Biomedical Imaging (ISBI) grand challenge for dental X-ray analysis conducted by IEEE. Furthermore, the proposed CephXNet is also tested with private clinical dataset comprised of 100 cephalograms collected from Solanki Dental Care Clinic in Sharjah, United Arab Emirates. The proposed CephXNet model obtained a classification accuracy of 90.08% and an average SDR of 73.94% in the 2mm precision range. The experimental outcomes show proposed CephXNet model is efficient and has great potential to deploy in clinical practice.

Keywords