Pneumonia classification: A limited data approach for global understanding

Abstract

As the human race has advanced, so too have the ailments that afflict it. Diseases such as pneumonia, once considered to be basic flu or allergies, have evolved into more severe forms, including SARs and COVID-19, presenting significant risks to people worldwide. In our study, we focused on categorizing pneumonia-related inflammation in chest X-rays (CXR) using a relatively small dataset. Our approach was to encompass a comprehensive view, addressing every potential area of inflammation in the CXR. We employed enhanced class activation maps (mCAM) to meet the clinical criteria for classification rationale. Our model incorporates capsule network clusters (CNsC), which aids in learning different aspects such as geometry, orientation, and position of the inflammation seen in the CXR. Our Capsule Network Clusters (CNsC) rapidly interpret various perspectives in a single CXR without needing image augmentation, a common necessity in existing detection models. This approach significantly cuts down on training and evaluation durations. We conducted thorough testing using the RSNA pneumonia dataset of CXR images, achieving accuracy and recall rates as high as 98.3% and 99.5% in our conclusive tests. Additionally, we observed encouraging outcomes when applying our trained model to standard X-ray images obtained from medical clinics.

Keywords

• Capsule networks clusters
• Deep learning
• Dicom
• Pneumonia
• X-rays