<![CDATA[Early detection of Autism Spectrum Disorder (ASD) is a crucial challenge in child development interventions because conventional screening methods are often subjective and prone to assessor bias. This study proposes an objective solution in the form of a deep learning approach for automatic ASD classification using facial landmark representations that focus exclusively on the eye and forehead areas. The selection of these areas is based on the eye avoidance hypothesis, which states that these regions contain very rich diagnostic information and behavioral biomarkers related to the ASD phenotype. The pre-processing stage involves isolating the eye and forehead areas using Dlib 68-landmark detection to eliminate background visual noise, followed by detailed topological visualization using MediaPipe Face Mesh with 478 landmark points as the model input. The Convolutional Neural Network (CNN) architecture used is the VGG19 model modified with transfer learning techniques and the addition of Dropout layers to improve efficiency and prevent overfitting. The model was trained on a primary dataset of 1,238 images collected under controlled conditions from children in Banda Aceh. The test results showed very promising performance with an overall accuracy of 94.35%. Specifically, the model achieved a recall (sensitivity) of 95.24%, a precision of 93.75%, and an AUC score of 0.9831. This high sensitivity is crucial in a medical context to minimize the risk of misdetection of positive cases. These results demonstrate that landmark visualization in the eye and forehead areas with the VGG19 model is a highly effective, accurate, and practical method for serving as an economical early screening tool for ASD.]]>
Copyrights © 2026
