Development of a landmark detection algorithm in resting state frontal face images using deep learning techniques

Objective: Orthodontic treatments can lead to significant changes in facial appearance. Therefore, a detailed evaluation of the face is important both at the diagnostic stage and in treatment planning. Traditionally conducted through anthropometric methods, these evaluations are now performed more practically using photographs. To enable a quantitative assessment, it is necessary to automatically identify anatomical reference points on the face with software support. This study investigates the capability of deep learning methods to accurately identify these reference points on facial photographs. Materials and Methods: While developing the deep learning model, 1000 facial photographs were annotated. The study employed the FarNet (Feature Aggregation and Refinement Network) architecture and the ResNet101 backbone network. The model was trained with a training dataset of 900 photographs and evaluated with a test dataset of 100 photographs. The performance of the developed deep learning model was assessed by calculating the mean radial error, standard deviation, and success detection rates. Results: The mean radial error of the developed deep learning model was found to be 2.02 ± 2.47 mm. The model achieved success detection rates of 70.41%, 77.33%, 81.94%, and 88.89% at intervals of 2.0 mm, 2.5 mm, 3.0 mm, and 4.0 mm, respectively. Among the 39 landmarks, the pupil right point had the lowest error value with a mean radial error of 0.58 ± 0.72 mm. The pupil right point achieved the highest success detection rates at intervals of 2.0 mm, 2.5 mm, 3.0 mm, and 4.0 mm. Conclusion: It has been observed that deep learning methods exhibit sufficient performance in identifying anatomical reference points.