ECAAF-CenterNet: Efficient Channel Attention and Adaptive Fusion for Cephalometric Landmark Detection

In response to the challenges of complex local structures and significant differences in scale distribution in key point detection of lateral cranial films, this paper proposes an automatic key point detection method for lateral cranial films based on the improved CenterNet, ECAAF-CenterNet. Based on the CenterNet anchor-free detection framework, DLA34 is selected as the backbone network. Combined with the characteristics of dense distribution of key points and obvious scale differences in head images, the network structure is improved from the two aspects of feature enhancement and feature fusion. Introduce the Efficient Channel Attention Mechanism (ECA) in the backbone residual module to enhance the network's expression ability for key channel features. Meanwhile, in the multi-scale feature fusion stage, an adaptive weighted cross-scale feature fusion mechanism AF is designed to perform dynamic weighted fusion on features at different levels. Through experimental verification on the private dataset of 1603 lateral cranial films, the MRE reached 1.54mm, and the SDR could reach 94.51% at the 4.0mm threshold, which was superior to the baseline model.