Deteksi Anomali pada Citra X-ray Dada Menggunakan Variational Autoencoder dengan Skema Sequential Hyperparameter Optimization

Diki Aulio Fransisko (Universitas Islam Negeri Sultan Syarif Kasim, Riau)
Febi Yanto (Universitas Islam Negeri Sultan Syarif Kasim, Riau)
Benny Sukma Negara (Universitas Islam Negeri Sultan Syarif Kasim, Riau)
Siti Ramadhani (Universitas Islam Negeri Sultan Syarif Kasim, Riau)
Reski Mai Candra (Universitas Islam Negeri Sultan Syarif Kasim, Riau)

TIN: TERAPAN INFORMATIKA NUSANTARA
Vol 7 No 3 (2026): August 2026

The limited availability of labeled medical images remains a major challenge in developing reliable deep learning-based disease detection systems. Conventional classification approaches generally require a large amount of abnormal data, whereas medical image annotation is time-consuming, costly, and highly dependent on radiological expertise. This study proposes an unsupervised anomaly detection model for chest X-ray images using a Variational Autoencoder (VAE), in which only normal images are utilized during the training process. The experiments were conducted on the COVID-19-Pneumonia-Normal Chest X-ray Images dataset, consisting of 5,228 images categorized into normal, pneumonia, and COVID-19 classes. The proposed framework includes image preprocessing, baseline VAE construction, sequential hyperparameter optimization, Beta-VAE implementation, and model evaluation using Accuracy, Precision, Recall, F1-Score, and the Area Under the Receiver Operating Characteristic Curve (AUROC). Experimental results demonstrate that the optimized model outperformed the baseline model, achieving an Accuracy of 97.50%, Precision of 96.32%, Recall of 100%, F1-Score of 98.12%, and an AUROC of 0.9999. These findings indicate that hyperparameter optimization and appropriate β coefficient selection improve latent representation learning, leading to more effective discrimination between normal and abnormal chest X-ray images. Therefore, the proposed approach has the potential to serve as an artificial intelligence-based early screening tool for chest radiograph analysis, particularly in scenarios where labeled medical data are limited.