Automated Cephalometric Landmark Detection: A Novel Software Model Compared with Manual Annotation Method

AI-based automated cephalometric landmark detection streamlines orthodontic diagnosis and treatment planning, providing accurate, efficient, and reliable results. Benefits include saving time, minimizing subjectivity, improving precision, and facilitating continuous improvement. However, they should complement clinician expertise, ensuring qualified orthodontists make the final diagnosis and treatment plan. To propose a method that automatically detects cephalometric landmarks on the X-ray images and compare these values with the manual annotation method. A dataset of 600 X-ray images, each containing 19 landmarks, was collected. Two orthodontists manually marked the 19 landmarks in 300 cephalograms and their coordinates were automatically extracted. The dataset was cleaned for errors, and a pre-trained CNN model with an EfficientNetB7 backbone was used for landmark detection. The model was trained on 80% of the dataset and tested on the remaining 20%. The two-step method included ROI extraction and landmark detection. The RMSE score was used to evaluate inter-examiner reliability and the R2 score was used to compare manual and automated models. Model landmark locations were compared to the manual method. The mean deviation of the predicted landmarks from the actual landmarks was calculated using RMSE, and the model showed acceptable accuracy compared to manual annotation. EfficientNetB7 was found to have detection accuracies similar to the manual annotation method. For landmarks like Porion, articulare, and soft tissue pogonion, the model outperformed the human annotation method and provides a consistent better result, and for points like Point A, pogonion, gnathion, and menton, the manual methods show more accurate results. The study introduced an automated approach using deep learning to predict landmark locations, and the results demonstrate its accuracy in comparison with the manual annotation method. This approach effectively detects cephalometric landmarks, suggesting its potential for clinical use with orthodontist’s supervision.