Saved in:
| Main Author: | |
|---|---|
| Format: | Recurso digital |
| Language: | |
| Published: |
Zenodo
2026
|
| Online Access: | https://doi.org/10.5281/zenodo.20095517 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Table of Contents:
- <p>Detecting Autism Spectrum Disorder (ASD) at an early stage is critically important for child development <br>and therapeutic intervention. However, building reliable artificial intelligence models for ASD screening <br>is hindered by strict data privacy regulations, limited availability of labeled facial image datasets, and <br>significant variations in facial characteristics across different regions and cultures. In this work, we <br>propose a federated learning (FL) framework for ASD detection using facial images, wherein raw data <br>remains locally stored at participating clinical centers and only trained model parameters are shared with <br>a central aggregation server. Each client performs standardized preprocessing including face extraction <br>via MediaPipe, removal of low-quality and duplicate images, and histogram-based normalization to <br>mitigate lighting and contrast variations. We systematically compare three convolutional neural network <br>(CNN) architectures—MobileNetV2, EfficientNetB0, and EfficientNetB4—trained in a federated setting <br>using the Federated Averaging (FedAvg) algorithm. EfficientNetB4 achieved the highest global model <br>accuracy of 91.7% and an F1-score of 90.3% after 20 federation rounds. The results demonstrate that <br>combining federated learning with careful client-side preprocessing significantly improves detection <br>accuracy while fully preserving data privacy. The proposed framework supports cross-cultural <br>applicability and is suitable for deployment in real-world clinical decision-support systems. </p>