Deep Learning for Assessment of Cardiac Chamber Enlargement on Anteroposterior Chest Radiographs

Area of Science:

• Artificial Intelligence in Medical Imaging
• Cardiovascular Diagnostics
• Radiology and Medical Imaging

Background:

• Cardiac chamber enlargement (CCE) is a significant indicator of cardiovascular disease.
• Accurate identification of CCE is crucial for patient management.
• Current methods for CCE detection on chest radiographs have limitations.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) algorithm for detecting CCE on anteroposterior (AP) chest radiographs.
• To utilize same-day transthoracic echocardiography (TTE) as the reference standard for algorithm training and validation.
• To compare the DL model's performance against cardiothoracic ratio (CTR) measurements and expert radiologist assessments.

Main Methods:

• A pretrained EfficientNet-B6 DL model was adapted to predict CCE from AP chest radiographs.
• A dataset of 6467 AP chest radiographs, acquired within 24 hours of TTE, was retrospectively collected and split into training, validation, and test sets.
• Model performance was evaluated using metrics like AUC, accuracy, sensitivity, and specificity, and compared to manual CTR and radiologist interpretations.

Main Results:

• The DL model achieved an AUC of 0.83 in the test set for binary CCE classification.
• The model demonstrated superior performance compared to manual CTR measurements (AUC 0.83 vs 0.74) and cardiothoracic radiologists (accuracy 77% vs 60.5%-67.5%).
• The algorithm showed robust performance with an accuracy of 76% and sensitivity of 81% in the test set.

Conclusions:

• A DL algorithm can effectively identify cardiac chamber enlargement on AP chest radiographs.
• The developed DL model outperforms conventional CTR measurements and expert radiologist assessments in a controlled setting.
• This study presents a proof-of-concept for AI-driven analysis of radiographic images for improved cardiac diagnostics.