Deep Learning-Based Image Analysis for the Prediction of Autism Spectrum Disorder

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition that affects communication, social interaction, and behaviour. Its early diagnosis remains challenging due to variability in symptom presentation, overlap with other developmental conditions, and reliance on subjective behavioural assessments. Conventional methods often delay intervention, which can limit opportunities for improving longterm developmental outcomes. To address this gap, the present study explores an image-based computational framework for supporting ASD prediction. A publicly available dataset containing facial images of autistic and non-autistic individuals was utilised for experimentation. The images were pre-processed through resizing, normalization, and augmentation techniques to enhance quality and increase robustness. The study employed a DenseNet201 model with modifications tailored for binary classification of ASD and non-ASD categories. Transfer learning was integrated to leverage prior knowledge, while dropout and batch normalization layers were applied to reduce overfitting and stabilize training. The model was trained and validated using structured splits of the dataset, with the Adamax optimizer and categorical cross-entropy loss function guiding parameter updates. A custom learning rate adjustment mechanism was employed to balance convergence and generalization. Performance evaluation was carried out through accuracy and loss plots across epochs, supported by separate testing on unseen samples. Experimental results demonstrated that the model achieved high training accuracy approaching near-perfect performance, with validation accuracy stabilizing around 90%. Despite fluctuations in validation loss, the framework showed strong potential for identifying autistic features from facial data. Visualization of sample predictions and error trends highlighted areas requiring refinement, particularly in enhancing consistency on unseen data. This study underscores the significance of computational screening tools in complementing conventional behavioural assessments. By extracting subtle facial patterns that may escape manual observation, such methods can serve as effective early indicators, particularly in resource-limited settings where specialist diagnosis is scarce. The findings also point to opportunities for future enhancement through larger datasets, additional regularization strategies, and integration with multimodal behavioural or neuroimaging data. In conclusion, the proposed approach provides an efficient and scalable foundation for supporting early ASD detection. While not intended as a standalone diagnostic tool, it offers valuable supplementary insights to clinicians, educators, and caregivers, ultimately contributing to improved developmental interventions and outcomes.