Deep transfer learning to quantify pleural effusion severity in chest X-rays.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Pulmonary medicine

Background:

• Pleural effusion detection in chest radiography is vital for managing chronic obstructive pulmonary disease (COPD).
• Accurate quantification of pleural effusion severity is essential for effective treatment planning.
• Deep learning offers a promising approach for automating medical image analysis.

Purpose of the Study:

• To develop and evaluate a deep learning model for quantifying pleural effusion severity in chest radiographs of COPD patients.
• To utilize the MIMIC-CXR database for training and validating the model.
• To assess the model's accuracy and interpretability in grading pleural effusion.

Main Methods:

• Utilized the MIMIC-CXR dataset, classifying patients with and without COPD.
• Extracted pleural effusion severity labels from radiology reports, categorizing into four grades.
• Trained and optimized various deep learning models (Resnet, DenseNet, GoogleNET) using different data processing and loss functions.
• Evaluated model performance using ROC curves, AUC, and confusion matrices; employed Grad-CAM for interpretation.

Main Results:

• The dataset comprised 15,620 chest X-rays from 8,533 patients, with varying degrees of pleural effusion.
• The optimized deep learning model achieved a highest accuracy rate of 73.07%.
• Micro-average AUCs for testing and validation cohorts were 0.89 and 0.90, respectively, with macro-average AUCs of 0.86 and 0.89.

Conclusions:

• Deep transfer learning models can effectively grade the severity of pleural effusion in chest radiographs.
• The developed model demonstrates potential for assisting clinicians in diagnosing and managing pleural effusion in COPD patients.
• Further refinement of deep learning models can enhance diagnostic accuracy in pulmonary imaging.