Computer-aided detection of interstitial abnormalities in chest radiographs using a reference standard based on...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Interstitial lung disease (ILD) presents subtle, ill-defined lesions on chest radiographs, challenging even expert radiologists.
• Accurate localization of these lesions is crucial for diagnosis and monitoring.

Purpose of the Study:

• To develop and evaluate a computer-aided detection (CAD) system for localizing interstitial lesions in chest radiographs.
• To propose a novel semiautomatic method for creating a reference standard using CT scans for CAD system training and evaluation.

Main Methods:

• A two-class supervised pattern classification approach analyzes lung fields pixel-wise using a multiscale Gaussian filter bank.
• A semiautomatic method transforms lesion outlines from CT scans to chest X-rays for ground truth generation.
• System performance was evaluated using Receiver Operating Characteristic (ROC) analysis.

Main Results:

• The CAD system achieved an area under the ROC curve (A(z)) of 0.78 with linear discriminant and support vector machine classifiers.
• The system demonstrated high performance (A(z) = 0.90) in differentiating obviously abnormal from normal pixels.
• System performance (A(z) = 0.85) was comparable to human observers (A(z) = 0.88) when excluding perihilar regions.

Conclusions:

• The developed CAD system shows promise for detecting interstitial lesions in chest radiographs.
• The semiautomatic reference standard generation method facilitates robust training and evaluation of CAD systems.
• The system's performance suggests it can be a valuable tool to assist radiologists in diagnosing ILD.