A critical review of artificial intelligence based techniques for automatic prediction of cephalometric landmarks

Abstract

Automatic cephalometric landmark detection has emerged as a pivotal area of research that combines medical imaging, computer vision, and orthodontics. The identification of cephalometric landmarks is of utmost importance in the field of orthodontics, as it contributes significantly to the process of diagnosing and planning treatments, as well as conducting research on craniofacial aspects. This practice holds the potential to improve clinical decision-making and ultimately increase the outcomes for patients. This work explores a wide range of strategies, encompassing both traditional edge-based methods and advanced deep learning approaches. The study leveraged various academic publication databases like IEEEXplore, ScienceDirect, arXiv, Springer and PubMed to thoroughly search for articles related to automatic cephalometric landmark detection. Additionally, other pertinent publications were acquired from credible sources like Google Scholar and Wiley databases. Screening the articles relied on three selection criteria: (a) publication titles, abstracts, literature reviews, (b) cephalometric radiograph datasets suitable for 2D landmarking, and (c) studies conducted over different time periods were employed to gain a comprehensive understanding of the evolution of methodologies used in landmark prediction to identify the most relevant papers for this review. The initial electronic database search identified 268 papers on landmark detection. A total of 118 publications were selected and incorporated in the present study after a meticulous screening process. Performance analysis was conducted on studies that reported Successful Detection Rates (SDRs) within different clinically accepted precision ranges, Mean Radial Error (MRE) with Standard Deviation (SD) between manually annotated and automated landmarks as outcomes. Bar graphs and custom combination plots were utilized to analyse the correlations among different methodologies employed and their evaluation metrics outcomes. The performance comparison results indicate that Deep Learning techniques showed superior accuracy in automating 2D cephalometric landmarks compared to other conventional and Machine Learning approaches. Recently, more advanced Deep Learning algorithms have been developed to improve the accuracy of automatic landmark prediction.