Electronic Theses and Dissertation

Seiring perkembangan kecerdasan buatan, khususnya deep learning, analisis citra
wajah menjadi pendekatan potensial untuk mendukung skrining Down Syndrome
(DS) secara otomatis. Penelitian ini berfokus pada analisis pengaruh strategi
pembagian dataset (data splitting) terhadap kinerja model klasifikasi citra wajah
DS dan Non-DS berbasis arsitektur EfficientNet-B0. Metode yang digunakan
meliputi tahapan pre-processing data seperti filtering kualitas citra, resizing,
normalisasi, serta augmentasi menggunakan Gaussian noise. Dataset yang
digunakan berasal dari platform Roboflow dengan total 2.620 citra setelah proses
seleksi. Penelitian ini menerapkan dua skenario data splitting, yaitu 70:20:10 dan
80:10:10 untuk data pelatihan, validasi, dan pengujian. Model dilatih menggunakan
parameter yang sama pada kedua skenario untuk memastikan perbandingan yang
adil, serta dievaluasi menggunakan metrik accuracy, precision, recall, dan F1
score yang divisualisasikan melalui confusion matrix. Hasil penelitian
menunjukkan bahwa skenario 70:20:10 menghasilkan performa yang lebih baik
dengan akurasi sebesar 93,56%, dibandingkan dengan skenario 80:10:10 yang
memperoleh akurasi sebesar 88,55%. Selain itu, skenario 70:20:10 juga
menunjukkan keseimbangan yang lebih baik pada nilai precision, recall, dan F1
score, serta tingkat kesalahan klasifikasi yang lebih rendah berdasarkan confusion
matrix. Berdasarkan hasil tersebut, dapat disimpulkan bahwa proporsi data splitting
memiliki pengaruh signifikan terhadap kinerja dan kemampuan generalisasi model.
Pembagian data yang seimbang antara data pelatihan dan validasi terbukti lebih
optimal dibandingkan peningkatan jumlah data pelatihan semata. Penelitian ini
memberikan kontribusi dalam menentukan strategi data splitting yang efektif untuk
meningkatkan performa model deep learning pada klasifikasi citra wajah DS.

With the advancement of artificial intelligence, particularly deep learning, facial image analysis has become a promising approach to support automatic screening of Down Syndrome (DS). This study focuses on analyzing the effect of dataset splitting strategies on the performance of a facial image classification model for DS and Non-DS using the EfficientNet-B0 architecture. The proposed method includes several stages, namely data pre-processing such as image quality filtering, resizing, normalization, and augmentation using Gaussian noise. The dataset was obtained from the Roboflow platform, consisting of 2,620 facial images after the selection process. Two data splitting scenarios were implemented, namely 70:20:10 and 80:10:10 for training, validation, and testing datasets. The model was trained using identical hyperparameters in both scenarios to ensure a fair comparison. Model performance was evaluated using accuracy, precision, recall, and F1-score metrics, which were further analyzed using a confusion matrix. The results show that the 70:20:10 data splitting scenario achieved better performance with an accuracy of 93.56%, compared to the 80:10:10 scenario which obtained an accuracy of 88.55%. In addition, the 70:20:10 configuration demonstrated more balanced precision, recall, and F1-score values, as well as lower misclassification rates based on the confusion matrix. It can be concluded that the proportion of data splitting significantly affects model performance and generalization ability. A balanced distribution between training and validation data is more effective than simply increasing the training data size. This study contributes to identifying an optimal data splitting strategy to improve deep learning model performance in Down Syndrome facial image classification.