Electronic Theses and Dissertation

Autism Spectrum Disorder (ASD) merupakan kondisi perkembangan saraf yang memengaruhi komunikasi sosial, interaksi, perilaku, dan respons sensorik anak. Diagnosis ASD umumnya memerlukan asesmen klinis komprehensif oleh tenaga profesional, sehingga membutuhkan waktu, biaya, dan akses layanan yang belum merata. Kondisi ini dapat menyebabkan keterlambatan intervensi dini. Oleh karena itu, diperlukan sistem pendukung skrining awal yang cepat, non-invasif, mudah diakses, dan tidak dimaksudkan untuk menggantikan diagnosis klinis. Penelitian ini mengembangkan sistem klasifikasi citra wajah anak dengan ASD dan anak non autis berbasis deep learning pada dataset primer anak Indonesia. Penelitian dilakukan dalam tiga tahap. Tahap pertama membangun Dataset Primer A yang terdiri atas 1.050 citra wajah dari 70 anak, yaitu 35 anak dengan ASD dan 35 anak non autis, dengan tiga ekspresi wajah: datar, tersenyum, dan sedih. Data dikumpulkan melalui prosedur akuisisi terstandardisasi dan persetujuan etik institusional, kemudian diproses melalui cropping, resizing 224×224 piksel, dan normalisasi. Pada tahap ini dikembangkan model baseline menggunakan CNN custom layer. Tahap kedua mengevaluasi strategi hard decision fusion berbasis kombinasi pairwise enam arsitektur CNN, yaitu ResNet-101, DenseNet-201, XceptionNet, ShuffleNet, MobileNetV2, dan EfficientNet-B0. Seluruh 15 kombinasi pasangan model dianalisis untuk memperoleh kombinasi terbaik berdasarkan akurasi, presisi, recall, F1-score, sensitivitas, dan efisiensi implementasi. Tahap ketiga melakukan perbandingan antara CNN dan Vision Transformer menggunakan Dataset Primer B yang diperluas menjadi 1.380 citra wajah. Model yang diuji meliputi ResNet50, DenseNet121, ViT-Tiny, dan DeiT-Tiny dengan dua strategi augmentasi, yaitu standard dan advanced augmentation, serta validasi statistik menggunakan paired t-test dan Cohen’s d effect size. Hasil penelitian menunjukkan bahwa CNN custom layer mencapai akurasi 90%, presisi 92%, recall 88%, dan F1-score 90%. Strategi hard decision fusion menunjukkan potensi peningkatan robustisitas dibandingkan model tunggal. Pada tahap analisis perbandingan, ResNet50 memperoleh performa terbaik dengan akurasi 99,28% dan AUC 0,9992, menunjukkan bahwa CNN tetap sangat kompetitif pada dataset primer berskala terbatas. Kontribusi utama penelitian ini adalah pembangunan dataset primer citra wajah anak Indonesia, pengembangan baseline CNN, evaluasi decision fusion pairwise enam CNN, perbandingan CNN dan Vision Transformer, serta penerapan evaluasi multimetrik dan validasi statistik sebagai fondasi pengembangan sistem skrining awal ASD berbasis AI.

Kata kunci: Autism Spectrum Disorder, citra wajah, Convolutional Neural Network, Vision Transformer, decision fusion, deep learning, skrining ASD.

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition that affects children’s social communication, interaction, behavior, and sensory responses. ASD diagnosis generally requires comprehensive clinical assessment by professionals; therefore, it often demands considerable time, cost, and access to specialized services that are not evenly available. This condition may lead to delayed early intervention. Accordingly, a fast, non-invasive, accessible early screening support system is needed, while not being intended to replace clinical diagnosis. This study develops a deep learning-based facial image classification system for children with ASD and Non Autis children using a primary dataset of Indonesian children. The study was conducted in three stages. The first stage developed Primary Dataset A, consisting of 1,050 facial images from 70 children, including 35 children with ASD and 35 Non Autis children, with three facial expressions: neutral, smiling, and sad. The data were collected through a standardized acquisition procedure with institutional ethical approval, and were then processed through cropping, resizing to 224×224 pixels, and normalization. At this stage, a baseline model was developed using a custom-layer CNN. The second stage evaluated a hard decision fusion strategy based on pairwise combinations of six CNN architectures, namely ResNet-101, DenseNet-201, XceptionNet, ShuffleNet, MobileNetV2, and EfficientNet-B0. All 15 pairwise model combinations were analyzed to identify the best combinations based on accuracy, precision, recall, F1-score, sensitivity, and implementation efficiency. The third stage performed a comparative analysis between CNN and Vision Transformer using Primary Dataset B, which was expanded to 1,380 facial images. The evaluated models included ResNet50 and DenseNet121 as CNN representatives, and ViT-Tiny and DeiT-Tiny as Vision Transformer representatives. The evaluation was conducted using two augmentation strategies, namely standard augmentation and advanced augmentation, and was supported by statistical validation using paired t-test and Cohen’s d effect size. The results showed that the custom-layer CNN achieved 90% accuracy, 92% precision, 88% recall, and 90% F1-score. The hard decision fusion strategy demonstrated the potential to improve classification robustness compared with single models. In the comparative analysis stage, ResNet50 achieved the best performance, with an accuracy of 99.28% and an AUC of 0.9992, indicating that CNN remains highly competitive on a limited-scale primary dataset. The main contributions of this study include the development of a primary facial image dataset of Indonesian children, the establishment of a CNN baseline model, the evaluation of pairwise decision fusion across six CNN architectures, the comparison between CNN and Vision Transformer, and the application of multimetrics evaluation and statistical validation as a foundation for developing an AI-based early ASD screening system. Keywords: Autism Spectrum Disorder, facial image, Convolutional Neural Network, Vision Transformer, decision fusion, deep learning, ASD screening.