Electronic Theses and Dissertation

Down syndrome (ds) merupakan kelainan genetik akibat trisomi kromosom 21 yang dapat dikenali melalui karakteristik morfologi wajah. pendekatan analisis citra wajah menjadi metode non-invasif yang potensial untuk mendukung deteksi dini ds. seiring perkembangan teknologi deep learning, ekstraksi fitur visual dari citra wajah dapat dilakukan secara otomatis dan akurat. oleh karena itu, penelitian ini bertujuan untuk membandingkan performa dua arsitektur lightweight convolutional neural network, yaitu efficientnet-b1 dan mobilenetv3-large, dalam mengklasifikasi citra wajah anak dengan ds dan non-ds. dataset yang digunakan terdiri dari 6.332 citra wajah yang diperoleh dari platform roboflow, kemudian melalui proses filtering dan preprocessing sehingga diperoleh 2.620 citra berkualitas tinggi. tahap preprocessing meliputi image enhancement menggunakan gaussian noise, image sharpening, dan contrast adjustment, serta resizing sesuai kebutuhan masing-masing model. dataset dibagi menjadi 70% data latih, 20% data validasi, dan 10% data uji. evaluasi performa dilakukan menggunakan metrik akurasi, precision, recall, dan f1-score, serta dianalisis menggunakan confusion matrix. selain itu, metode grad-cam digunakan untuk meningkatkan interpretabilitas model. hasil penelitian menunjukkan bahwa efficientnet-b1 menghasilkan performa yang lebih baik dengan akurasi sebesar 91.3% dibandingkan mobilenetv3-large sebesar 86.4%. selain itu, efficientnet b1 juga menunjukkan stabilitas yang lebih baik serta visualisasi grad-cam yang lebih terfokus pada area wajah yang relevan. dengan demikian, efficientnet-b1 dapat menjadi pilihan yang lebih efektif dalam pengembangan sistem deteksi dini ds berbasis citra wajah yang akurat, efisien, dan mudah diinterpretasikan.

Down Syndrome (DS) is a genetic disorder caused by trisomy of chromosome 21, which can be identified through characteristic facial morphology. Facial image analysis has emerged as a promising non-invasive approach to support early detection of DS. With the advancement of deep learning technology, visual feature extraction from facial images can be performed automatically and accurately. Therefore, this study aims to compare the performance of two lightweight convolutional neural network architectures, namely EfficientNet-B1 and MobileNetV3-Large, in classifying facial images of children with and without DS. The dataset consists of 6,332 facial images obtained from the Roboflow platform, which were then processed through filtering and preprocessing stages, resulting in 2,620 high-quality images. The preprocessing steps include image enhancement using Gaussian noise, image sharpening, and contrast adjustment, as well as resizing according to each model’s input requirements. The dataset was divided into 70% training, 20% validation, and 10% testing data. Model performance was evaluated using accuracy, precision, recall, and F1-score, and further analyzed using a confusion matrix. In addition, Grad-CAM was applied to enhance model interpretability by highlighting the most relevant facial regions for classification. The experimental results show that EfficientNet-B1 achieved better performance with an accuracy of 91.3% compared to MobileNetV3-Large at 86.4%. Furthermore, EfficientNet-B1 demonstrated more stable learning behavior and produced more focused Grad-CAM visualizations on relevant facial regions. In conclusion, EfficientNet-B1 is more effective for developing an accurate, efficient, and interpretable facial image-based system for early detection of DS.