Integrating Machine Learning Techniques to Improve Pneumonia Diagnostics by Analyzing Chest X-ray Scans

Publication Date

1-1-2024

Document Type

Conference Proceeding

Publication Title

2024 IEEE Integrated STEM Education Conference, ISEC 2024

DOI

10.1109/ISEC61299.2024.10665026

Abstract

Pneumonia is the fourth most common cause of mortality overall, and the COVID-19 pandemic has only worsened these numbers. With the emergence of highly efficient AI/ML tools, pneumonia diagnostics can be improved in terms of time, accuracy, and cost, paving the way for timely treatments. In the current work, 998 chest X-rays were obtained from the RSNA International COVID-19 Open Radiology Database (RICORD). Each X-ray image was evaluated by three radiologists based on appearance (typical, indeterminate, atypical, or negative for pneumonia) and airspace disease grading (mild, moderate, or severe). The convolutional neural network (CNN) algorithm was employed on four different variations of the dataset described above - the diagnoses of radiologist #1, radiologist #2, and radiologist #3 as well as a three-timed-duplicated set including each of the three diagnoses based on a single chest X-ray scan as a separate entry. The same CNN model achieved testing accuracies of 44.39%, 18.93%, 20.00%, and 27.93% respectively. As expected, the impact of subjectivity can be identified in terms of low model accuracies. Poor to moderate model performance across all four classification tasks indicates the problem that non-objective evaluations of chest X-rays, specifically variations in the diagnostic analysis of ten similar scenarios, play in medical decisions.

Keywords

chest X-rays, machine learning, pneumonia

Department

Mechanical Engineering

Link to Full Text

Share

COinS