<![CDATA[Anomaly detection in chest X-ray images remains a challenge in medical imaging, as supervised learning approaches require large amounts of labeled data that are difficult and costly to annotate. This study proposes an unsupervised learning anomaly detection system that integrates a pretrained DenseNet121 as a feature extractor with a Denoising Autoencoder (DAE), so that training requires only normal images without anomaly annotations. The model was trained on normal images and tested on COVID-19 and pneumonia images to evaluate its anomaly detection capability based on reconstruction error relative to an optimal threshold. Evaluation was conducted on the Covid19-Pneumonia-Normal Chest X-Ray Images dataset comprising 5,228 images, comparing the performance of DenseNet121 and ResNet50 as feature extractors across three latent dimension configurations. The DenseNet121 configuration with a latent dimension of 128 achieved the highest overall performance on most metrics, namely 91% accuracy, 90.75% sensitivity, 72% Macro F1-Score, and a validation loss (MSE) of 0.0952 on 3,608 test images, although its AUC (0.9526) and specificity (87.36%) were not consistently the highest among all tested configurations. These results demonstrate that using DenseNet121 as a feature extractor improves the DAE's ability to distinguish normal from anomalous lung images, suggesting its potential as an efficient preliminary screening approach under conditions of limited labeled data.]]>
