Electronic Theses and Dissertation

Autism Spectrum Disorder (ASD) merupakan gangguan perkembangan saraf yang
memerlukan deteksi dini untuk mendukung pemberian intervensi yang lebih efektif.
Perbedaan morfologi wajah pada anak dengan ASD membuka peluang pemanfaatan
citra wajah sebagai media skrining awal berbasis kecerdasan buatan. Penelitian ini
bertujuan mengembangkan metode klasifikasi wajah anak dengan ASD menggunakan
segmentasi fitur wajah berbasis facial landmark detection dan ekstraksi deep feature
berbasis Convolutional Neural Network (CNN), serta membandingkan performa
kombinasi backbone CNN dan metode klasifikasi. Tahapan penelitian meliputi deteksi
wajah, deteksi 68 titik facial landmark, segmentasi empat Region of Interest (ROI)
berupa mata kiri, mata kanan, hidung, dan mulut, ekstraksi deep feature menggunakan
ResNet50 dan MobileNetV2, feature fusion melalui konkatenasi, serta proses
klasifikasi menggunakan Extreme Gradient Boosting (XGBoost) dan Multilayer
Perceptron (MLP). Evaluasi dilakukan menggunakan metrik accuracy, precision,
recall, F1-score, Area Under Curve (AUC), serta 5-fold stratified cross validation.
Hasil penelitian menunjukkan bahwa kombinasi ResNet50 dan MLP memberikan
performa terbaik dengan accuracy 98,23%, precision 0,9851, recall 0,9851, F1-score
0,9851, dan AUC 0,9994 pada test set. Evaluasi cross validation juga menunjukkan
kemampuan generalisasi yang baik dengan accuracy rata-rata 97,58% ± 0,0097 tanpa
indikasi overfitting yang signifikan. Hasil tersebut menunjukkan bahwa pendekatan
segmentasi fitur wajah yang dipadukan dengan ekstraksi deep feature berbasis CNN
mampu menghasilkan representasi fitur yang efektif untuk membantu proses skrining
awal ASD berbasis citra wajah.

Autism Spectrum Disorder (ASD) is a neurodevelopmental disorder that requires early detection to support timely intervention and improve developmental outcomes. Differences in facial morphology among children with ASD provide an opportunity to utilize facial images as a non-invasive medium for artificial intelligence-based early screening. This study aims to develop a facial image classification method for children with ASD using facial feature segmentation based on facial landmark detection and Convolutional Neural Network (CNN)-based deep feature extraction, while comparing the performance of different CNN backbones and classification methods. The proposed framework consists of face detection, 68-point facial landmark detection, segmentation of four Regions of Interest (ROIs), namely the left eye, right eye, nose, and mouth, deep feature extraction using ResNet50 and MobileNetV2, feature fusion through feature concatenation, and classification using Extreme Gradient Boosting (XGBoost) and Multilayer Perceptron (MLP). Model performance was evaluated using accuracy, precision, recall, F1-score, Area Under the Curve (AUC), and 5-fold stratified cross validation. The experimental results demonstrate that the ResNet50- MLP combination achieved the best performance, obtaining an accuracy of 98.23%, precision of 0.9851, recall of 0.9851, F1-score of 0.9851, and an AUC of 0.9994 on the independent test set. Furthermore, the cross-validation results indicate strong generalization capability, with a mean accuracy of 97.58% ± 0.0097 and no significant indication of overfitting. These findings demonstrate that facial feature segmentation combined with CNN-based deep feature extraction provides an effective feature representation for supporting early ASD screening using facial images.