CephTransXnet: An attention enhanced feature fusion network leveraging neighborhood rough set approach for cephalometric landmark prediction

Abstract

The convergence of medical imaging, computer vision, and orthodontics has made automatic cephalometric landmark detection a pivotal area of research. Accurate cephalometric analysis is crucial in orthodontics, orthognathic and maxillofacial surgery for diagnosis, treatment planning, and monitoring craniofacial growth. In this research study, a multi-branch fused feature extraction network titled ${CephTransX}_{net}$ is proposed to automatically predict landmark coordinates from cephalometric radiographs. The initial sequential branch enhances discriminative local feature learning and feature extraction through parallel feature fusion by integrating Convolved Pooled Normalized (${CPN}_B$) and Gradient Optimized Multi-Path Bottleneck (${GMB}_B$) blocks with Channel and Spatial Attention (${CSAT}_M$) module. The Swin Transformer (${ST}_B)$ branch efficiently handles long-range dependencies and extracts global features in cephalometric radiographs. The multi-branch fused features along with features from skip connections of ${CPN}_B$ and ${GMB}_B$ blocks are concatenated using a Coordinate Attention module $\left({CoAT}_M\right)$ to captures the positional relationships between various landmark features. A Landmark Discriminative Deviation Factor $\left(LDDF\right)$ is determined by applying the Neighborhood Rough Set $\left(NRS\right)$ approach to analyse the surrounding features of each landmark by considering spatial relationships or similarity measures between the landmarks and neighboring regions. The Spatial Pyramid Pooling (${SPP}_L$) layer incorporated in the final phase of ${CephTransX}_{net}$ model extracts multi-scale features by pooling over sub-regions of varying sizes, enabling the network to capture both local and global context for precise cephalometric landmark identification. The ${CephTransX}_{net}$ framework achieved an average Successful Detection Rates $\left({SDR}_s\right)$ of 88.71 % and 79.05 % in 2 mm using the 2015 International Symposium on Biomedical Imaging (ISBI) grand challenge dental X-ray analysis dataset. The effectiveness of the ${CephTransX}_{net}$ model is evaluated using a private clinical dataset obtained from Solanki Dental Care Clinic in Sharjah, UAE, and attained an average ${SDR}_s$ of 74.38 % in 2 mm precision range.