Extended Template Matching method for Region of Interest Extraction in Cephalometric Landmarks Annotation

Abstract

Finding areas in the image where the subsequent processing of the features concentrates is known as Region of Interest (ROI) extraction. Utilizing ROI helps speed up processing by excluding irrelevant image regions. ROI extraction in biomedical landmark annotation problems is challenging as radiograph images have varying contrast and intensity levels. Cephalometric landmark annotation is a domain where ROI extraction plays a vital role in traditional machine learning and deep learning solutions. This work proposes a simple and feasible extension to the template matching method to extract the ROI from the cephalometric images. The exact ROI patch is located based on a combined metric calculated using the Normalized correlation coefficient measure and the distance measure. The algorithm is tested on publicly available cephalometric landmark annotation dataset. The experimental results show that the ROIs are extracted with an accuracy of 99.69%. Additionally, a reported average distance between the ROI patch center and the ground truth landmark is 3.96 mm. This demonstrates that the method can practically be used as an initial estimator, significantly improving the accuracy of landmark localization.