The mandible is crucial in orthodontic treatment, forensic identification, and clinical diagnosis. However, manually identifying mandibular landmarks is time-consuming and highly dependent on expert skill, necessitating a more reliable automated prediction method. Previous research has used linear and nonlinear regression methods, in which each model predicts a single landmark point, resulting in inefficiency. In addition, these methods only use the centroid of the binary image of the mandible as input. This research proposes a multi-output neural network model that is able to predict multiple mandibular landmark points simultaneously. The proposed neural network architecture in predicting mandibular landmark points is 11 neurons on the input layer, 22 neurons on the hidden layer, 32 neurons on the hidden layer, and 14 neurons on the output layer. The dataset consists of grayscale panoramic radiographs from the Dental and Oral Hospital, Faculty of Medicine, Universitas Airlangga, with 96 images for training and 6 for testing. Image features, including mean intensity, standard deviation, median, variance, skewness, kurtosis, entropy, contrast, homogeneity, energy, and correlation, were extracted and used as inputs to a multi-output neural network. The model predicted the reference points of the right condyle, left condyle, right coronoid, left coronoid, right gonion, left gonion, and menton. The results showed that the proposed model effectively predicted the reference points of the mandible, with the right condyle showing the highest accuracy. The highest prediction accuracy value, with a Successful Detection Rate (SDR) 12% and the Adam optimizer, was the right condyle point.
Nafiiyah et al. (2026) studied this question.