Pneumonia Detection with CheXNet: A Deep Learning Model Surpasses Radiologist Accuracy
Introduction
Pneumonia is a leading cause of hospitalization in the United States, with over 100 million adults admitted annually. Chest X-ray is the primary diagnostic tool for pneumonia, but detecting the disease through X-rays is a challenging task that requires radiologists’ expertise. In this work, we propose a deep learning model, CheXNet, that can automatically detect pneumonia from chest X-rays with higher accuracy than radiologists.
Background
The World Health Organization (WHO) has recognized chest X-ray as the best diagnostic tool for pneumonia (WHO, 2001). However, the accuracy of X-ray-based diagnosis relies heavily on radiologists’ professional competence. The Centers for Disease Control and Prevention (CDC) reported that in 2017, over 50,000 people died from pneumonia in the United States (CDC, 2017).
CheXNet Architecture and Training
CheXNet is a convolutional neural network (CNN) with 121 layers, designed to detect pneumonia from chest X-rays. The network takes a chest X-ray image as input and outputs a probability heatmap indicating the presence of pneumonia. We trained CheXNet on the ChestX-ray14 dataset, which contains 112,120 individual chest X-rays labeled with 14 different breast diseases, including pneumonia.
To optimize the network, we used intensive connections (dense connections) and batch normalization. We initialized the network weights randomly and used Adam’s optimization algorithm with standard parameters (β1 = 0.9 and β2 = 0.999). The learning rate was set to 0.01, and we trained the model with a batch size of 16.
Performance Comparison with Radiologists
We collected 420 chest X-rays and asked four practicing radiologists at Stanford University to annotate them. We then evaluated the performance of each radiologist and compared it with CheXNet’s performance. The results showed that CheXNet outperformed the radiologists in detecting pneumonia, with an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.788.
Model Explanation
To explain the network’s prediction, we used class activation mappings (CAMs) to generate a heatmap indicating the most indicative area of the image. The CAMs revealed that CheXNet focuses on specific features of the X-ray image to detect pneumonia.
Conclusion
CheXNet, a deep learning model, has surpassed the accuracy of radiologists in detecting pneumonia from chest X-rays. This achievement has significant implications for clinical care and epidemiological studies. CheXNet’s performance demonstrates the potential of deep learning in medical imaging and highlights the need for further research in this area.
Figure 1: CheXNet Architecture
[Insert Figure 1: CheXNet Architecture]
Figure 2: ROC Curve Comparison
[Insert Figure 2: ROC Curve Comparison]
References
CDC (2017). Deaths from pneumonia in the United States, 2017. Centers for Disease Control and Prevention.
WHO (2001). Diagnosis of pneumonia. World Health Organization.