Comparative Analysis of Regression and Classification Approaches for Landmark Localization in Cephalometric Analysis

Cephalometric landmark identification plays very important role in the process of cephalometric analysis. Manual annotation is challenging, time-consuming and subjective in nature. This study presents a comprehensive comparative examination of non-deep learning methods utilizing regression and classification techniques for automatically localizing cephalometric landmarks. The proposed framework leverages template matching and correlation values to efficiently pinpoint the initial potential landmark locations, thus streamlining the cephalometric landmark detection process. To accomplish this, the framework extracts neighboring local pixel intensity values as pixel descriptors and utilizes them to train a lightweight gradient boosting machine. In the classification approach, the center pixel is deemed the landmark point within a given patch, and patches are categorized as either landmark patches or non-landmark patches. On the other hand, the regression approach estimates the distance between the center of a patch and the ground truth pixel. The classification-based method demonstrates a significant 2% improvement in detection accuracy across public and private datasets, highlighting its superior performance over the regression-based approach.