Autism Spectrum Disorder Prediction from Facial Images Using Fine-Tuned Efficient Net B0–B7 Architectures

This research evaluates the effectiveness of the Efficient Net model series (B0–B7) in detecting Autism Spectrum Disorder using facial image data. The findings indicate that the ‎deeper models attain better accuracy and more balanced classification results than the ‎shallower models. EfficientNetB3 and B7 achieve the top accuracy of 0.99, exhibiting excellent precision, ‎recall, and F1-scores for both ASD and non-ASD categories, emphasizing their effectiveness in reducing false ‎positives and false negatives. EfficientNetB2 and B5 also reach competitive accuracies of 0.98 and ‎‎0.97, offering dependable options with marginally lower complexity. Conversely, EfficientNetB4 achieves the ‎lowest accuracy of 0.88 because of poor recall in the non-ASD category, indicating restricted generalization. ‎The results affirm that more profound Efficient Net models, especially B3, B5, B6, and B7, excel in feature ‎extraction and classification for ASD prediction, providing a trustworthy structure for the early ‎and precise detection of the disorder.