Improving efficiency of autism detection based on facial image landmarks

Abstract

Autism is a serious mental health problem with long-term effects on life. Therefore, early diagnosis is a topical issue for effective treatment. This study proposes a novel facial landmark transformation-based data augmentation method that allows for the generation of geometric transformations related to facial geometry. This method increases the generalizability and provides a perspective on the role of facial regions in autism detection. The proposed augmentation method ensures the generation of variants that are consistent with the facial image structure and the nature of the facial image. Next, conduct a comprehensive and comparative study with EfficientNet-B0, EfficientNet-B4, ResNet-18, ResNet-50, ResNet-101, MobileNet-V2, DenseNet-121 and DenseNet-201. Also analyze the model's attention over the main regions of the face that are related to facial landmarks. The results clearly show that the models trained with the proposed method outperform the default augmentation method. Specifically, when averaging the measures across the tested models, the results are 0.905417 for accuracy, 0.962133 for area under the curve (AUC), 0.9198 for precision, 0.888333 for recall, and 0.903678 for F1-score. Furthermore, when analyzing the gradient-weighted class activation mapping (Grad CAM) heatmaps, the high-value regions are clearly concentrated on the main areas of the face. Source code is published on GitLab platform.