Electronic Theses and Dissertation

Autism Spectrum Disorder (ASD) merupakan gangguan perkembangan saraf yang perlu dideteksi sejak dini untuk mendukung penanganan yang tepat. Penelitian ini bertujuan membandingkan performa tiga arsitektur EfficientNet, yaitu EfficientNetB0, EfficientNetB1, dan EfficientNetB2, dalam klasifikasi citra wajah anak ke dalam kategori autisme dan non autisme. Dataset yang digunakan terdiri dari 1.380 citra wajah, kemudian diproses melalui deteksi wajah menggunakan MTCNN dan augmentasi data (rotasi, flip, zoom, brightness adjustment, dan Gaussian noise). Model dilatih menggunakan metode transfer learning berbasis ImageNet dan fine tuning pada 10 lapisan akhir. Hasil menunjukkan bahwa EfficientNetB0 memberikan performa terbaik dengan akurasi pelatihan 99,83%, validasi 98,19%, dan pengujian 98,55%, serta nilai loss pengujian 4,08%. EfficientNetB1 mencapai akurasi pengujian 97,83% dengan loss 5,53%, sementara EfficientNetB2 juga memiliki akurasi 97,83% dengan nilai loss sebesar 4,22%. Ketiganya menunjukkan kinerja tinggi. Namun, EfficientNetB0 dinilai paling optimal karena efisien secara komputasi, stabil saat pelatihan, memiliki loss pengujian terkecil, dan risiko overfitting yang lebih rendah. Hasil ini mengindikasikan bahwa EfficientNetB0 layak digunakan sebagai solusi efektif untuk sistem klasifikasi wajah berbasis deep learning dalam mendukung skrining awal ASD secara praktis.

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder that needs to be detected early to support appropriate treatment. This study aims to compare the performance of three EfficientNet architectures, EfficientNetB0, EfficientNetB1, and EfficientNetB2, in classifying children's facial images into autism and non autism categories. The dataset consists of 1,380 facial images, which were processed through facial detection using MTCNN and data augmentation (rotation, flipping, zooming, brightness adjustment, and Gaussian noise). The model was trained using transfer learning based on ImageNet and fine-tuning on the last 10 layers. The results show that EfficientNetB0 provides the best performance with a training accuracy of 99.83%, validation accuracy of 98.19%, and testing accuracy of 98.55%, as well as a testing loss value of 4.08%. EfficientNetB1 achieved a test accuracy of 97.83% with a loss of 5.53%, while EfficientNetB2 also had an accuracy of 97.83% with a loss value of 4.22%. All three models demonstrate high performance, but EfficientNetB0 is considered the most optimal due to its computational efficiency, stability during training, smallest testing loss, and lower risk of overfitting. These results indicate that EfficientNetB0 is a viable solution for deep learning-based facial classification systems to support practical initial screening for ASD.