Clinical Equivalence of a CNN-Based Automated Soft Tissue Landmark Detection System on 2D Facial Images

Background/Objectives: The aim of this study was to evaluate and compare the accuracy, reliability, and time efficiency of a convolutional neural network (CNN)-based deep learning model with manual annotation in the identification of soft tissue landmarks on two-dimensional (2D) facial images for orthodontic applications. Materials and Methods: Three-dimensional (3D) facial scans were obtained from 100 participants (50 females, 50 males) aged 18-25 years using the Revopoint Pop2 3D Scanner. Frontal and profile 2D images were extracted from the 3D models. Manual landmark identification was performed by a single investigator using LabelMe software, marking 22 landmarks on frontal images and 15 landmarks on profile images. A novel CNN model was developed and trained on these manually annotated images. The model's automatic landmark identifications were compared with manual annotations in terms of positional error, identification time, and reproducibility. Results: The CNN model achieved a mean localization accuracy of 96.07%. The mean prediction error ranged from 2.3% to 4.5% across various anatomical points. Trichion, Menton, and Gonion points exhibited relatively higher error rates. The model significantly reduced the annotation time compared to manual identification (manual method: 237 s per image). Intra-observer reliability analysis demonstrated excellent agreement for manual landmarking (ICC: 0.85-0.95). The AI model provided consistent predictions for identical inputs. Conclusions: The deep learning-based model demonstrated comparable accuracy to manual landmark identification while significantly improving the annotation speed and reproducibility. These results suggest that CNN-based systems offer a promising alternative for clinical orthodontic analysis and digital workflow integration.