Evaluation Of Accuracy And Reliability Of Artificial Intelligence-Based Fully Automated And Semi-Automated Cephalometric Analysis Software In Comparison With Manual Cephalometric Analysis

Abstract

Introduction: Lateral cephalometric analysis stands as a pivotal diagnostic tool within the realm of Orthodontics, often regarded as the gold standard. Over time, several computerised programs have surfaced, aiming to streamline the process of digital cephalometric analysis, thereby offering efficiency gains and time savings. The current study sought to assess the diagnostic accuracy and reliability of four computerised cephalometric analysis programs. These encompassed three digital software solutions alongside an online artificial intelligence platform. The objective was to compare the digital approaches against the traditional method of manual cephalometric tracing.

Materials and methodology: This study analysed thirty pre-treatment lateral cephalometric radiographs of individuals undergoing orthodontic treatment. Four computer programs—Dolphin Imaging software, Nemoceph, and FACAD for semi-automated cephalometric analysis, and WebCeph for artificial intelligence-based cephalometric analysis—were evaluated and compared with manual cephalometric tracing. Measurements were taken for five skeletal, four dental, and one soft tissue parameters, including six angular and four linear measurements derived from Steiner, McNamara, and Tweed cephalometric analysis. Fifteen cephalometric radiographs were remeasured to assess the intra-examiner correlation coefficient. One-way ANOVA was used to compare different cephalometric points among the various tracing methods, and multiple comparisons were performed using the Post Hoc Tukey’s test.

Results: No significant difference was noted in SNA, SNB, ANB, FMA, IMPA, 1 to Pt A, 1 to N-A and nasolabial angle between the software. While 1 to A Pog and 1 to N-B revealed significant differences between the software. FACAD vs. NemoCeph showed a significant difference with a mean difference of -0.30 and -0.30 and a p-value of 0.03 and 0.02 for 1 to A Pog and 1 to N-B values respectively and this variability was noted between FACAD and NemoCeph.

Conclusion: Digital and AI-based software have shown accuracy comparable to manual tracing. Among the software, Dolphin Imaging proved to be the most reliable, followed by WebCeph and NemoCeph, with FACAD showing the most variability. The variability, which lacked clinical significance, was primarily due to challenges in identifying certain landmarks.