Capsule Network for Pneumonia Classification: Enhancing Robustness and Accuracy

Pneumonia is a severe global health threat, causing around 4 million deaths per year. Timely and accurate diagnosis is essential for effective treatment and reducing mortality rates. This study introduces a deep learning model called Capsule Networks (CapsNet) for classifying pneumonia, which can capture complex spatial relationships in chest X- ray (CXR) images. We utilized a dataset from Guangzhou Women and Children’s Medical Center, which was processed through SMOTE techniques before training the model. In order to mimic the distortions that may occur when images are transmitted in real-life scenarios, we applied rotations and flips to the images in the test set to compare and verify the robustness of CapsNet. The results showed that CapsNet achieved an accuracy rate of 89.58% and 87.02% before and after test set distortion, demonstrating good robustness. In terms of evaluation metrics, CapsNet outperformed other models like ResNet, Xception, and ViT.