M.S. AAI Capstone Chronicles 2024

CNN Lung Disease Classification

2

Introduction Neural networks excel at classification tasks due to their ability to learn and understand complex patterns. This capability can be applied to many fields, including medical imaging, where notably Convolutional Neural Networks (CNNs) have shown promise in detecting diseases from X-ray images (Varshni et al., 2019). This research paper aims to leverage CNNs in an attempt to accurately classify chest X-rays to multiple lung diseases. Three conditions are studied in this report, including Effusion, Atelectasis, and Infiltration, with “No Finding” included for a baseline comparison. By including “No Finding,” we establish a control group that represents healthy individuals, which improves our approach to distinguish between diseased and healthy cases. We will compare the results of several models including CNNs built from scratch, a CNN designed using a fine-tuned approach, and a pre-trained model. According to the Pan American Health Organization, chronic obstructive pulmonary disease (COPD) and lower respiratory infections are the fourth and fifth leading causes of death in North and South America (PAHO, 2019). Our goal is to study the early detection of acute respiratory conditions in an effort to curb the prevelance of chronic diseases that can lead to patient mortality. While the three conditions discussed in this paper are not chronic respiratory diseases themselves, they can contribute to or result from chronic respiratory diseases if not properly managed. By classifying these conditions through neural networks, the goal is for healthcare providers involved in lung disease diagnosing to work more effectively and reduce the global burden of respiratory illnesses. This paper serves as an outline for how this could be developed in a real-world setting on X-ray data. For this project, we utilized the National Institutes of

160