Optimizing convolutional neural networks for Chronic Obstructive Pulmonary Disease detection in clinical computed...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Pulmonary Medicine

Background:

• Chronic Obstructive Pulmonary Disease (COPD) diagnosis relies on identifying emphysema in lung computed tomography (CT) scans.
• Image window settings significantly impact the visibility of emphysema, affecting diagnostic accuracy.
• Convolutional Neural Networks (CNNs) show promise for automated medical image analysis.

Purpose of the Study:

• To optimize the binary detection of COPD based on emphysema presence using CNNs.
• To compare the efficacy of manually adjusted versus automated window-setting optimization (WSO) on CT images for COPD detection.
• To identify the most efficient CNN architecture for this task.

Main Methods:

• Retrospective analysis of 7194 contrast-enhanced CT images from 78 subjects (3597 COPD, 3597 controls).
• Exploration of manually clipped emphysema window settings and a full-range window setting.
• Implementation of automated WSO by integrating a customized layer into CNN models, including DenseNet.
• Evaluation of model performance using the area under the Receiver Operating Characteristics curve (AUC).

Main Results:

• DenseNet was identified as the most efficient CNN backbone, achieving a mean AUC of 0.80 without WSO.
• Manual adjustment of CT images to the emphysema window setting improved COPD detection by DenseNet, yielding a mean AUC of 0.86.
• Automated WSO with a customized DenseNet layer achieved a mean AUC of 0.82, learning optimal window settings near the emphysema range.

Conclusions:

• Manual window-setting optimization to the emphysema range significantly enhances COPD detection accuracy with DenseNet CNN models.
• Automated WSO offers a viable alternative, though manual optimization currently provides superior performance.
• Window-setting optimization is crucial for improving CNN-based COPD detection from CT data.