Electronic Theses and Dissertation

Autism Spectrum Disorder (ASD) merupakan gangguan perkembangan saraf yang memengaruhi interaksi sosial dan komunikasi pada anak. Deteksi dini terhadap ASD sangat penting untuk mendukung intervensi dan terapi yang tepat. Namun, diagnosis konvensional masih bergantung pada tenaga medis profesional dan membutuhkan waktu yang cukup lama. Penelitian ini bertujuan untuk mengembangkan sistem klasifikasi otomatis berbasis deep learning yang mampu membedakan wajah anak ASD dan normal secara objektif dan efisien menggunakan citra wajah. Dataset yang digunakan terdiri dari data primer yang dikumpulkan langsung dari subjek penelitian dengan total dataset 700 citra. Model dilatih menggunakan tiga arsitektur CNN, yaitu EfficientNet-B0, MobileNetV3, dan ShuffleNet-V2. Data latih diperluas melalui augmentasi geometrik dan dilatih menggunakan pendekatan early stopping untuk mencegah overfitting. Evaluasi dilakukan berdasarkan metrik akurasi, presisi, recall, dan F1-score. Hasil evaluasi menunjukkan bahwa arsitektur EfficientNet-B0 tanpa early stopping memberikan performa terbaik, dengan akurasi, presisi, recall, dan F1-score masing-masing sebesar 90%. Sementara itu, MobileNetV3 menunjukkan performa tinggi saat menggunakan early stopping dengan akurasi sebesar 88,57%. Perbandingan tersebut menunjukkan bahwa efektivitas strategi pelatihan sangat bergantung pada karakteristik model yang digunakan. Secara keseluruhan, sistem klasifikasi yang dikembangkan berhasil mendeteksi wajah anak ASD secara akurat dan efisien, serta berpotensi untuk digunakan sebagai alat bantu skrining awal di lingkungan dengan keterbatasan tenaga medis.

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder that affects social interaction and communication in children. Early detection of ASD is crucial for supporting appropriate intervention and therapy. However, conventional diagnosis still relies on medical professionals and requires a considerable amount of time. This study aims to develop an automated classification system based on deep learning that can objectively and efficiently distinguish between the faces of children with ASD and those without using facial images. The dataset used consists of primary data collected directly from research subjects, totaling 700 images. The model was trained using three CNN architectures: EfficientNet-B0, MobileNetV3, and ShuffleNet-V2. The training data was expanded through geometric augmentation and trained using the early stopping approach to prevent overfitting. Evaluation was conducted based on accuracy, precision, recall, and F1-score metrics. The evaluation results showed that the EfficientNet-B0 architecture without early stopping provided the best performance, with accuracy, precision, recall, and F1-score of 90% each. Meanwhile, MobileNetV3 showed high performance when using early stopping with an accuracy of 88.57%. This comparison shows that the effectiveness of the training strategy is highly dependent on the characteristics of the model used. Overall, the developed classification system successfully detected the faces of children with ASD accurately and efficiently and has the potential to be used as an initial screening tool in environments with limited medical personnel