CELM: An Ensemble Deep Learning Model for Early Cardiomegaly Diagnosis in Chest Radiography

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Cardiology

Background:

• Cardiomegaly, an enlarged heart, is a critical radiological sign of cardiovascular disease.
• Early detection of cardiomegaly is crucial for timely intervention and improved patient outcomes.
• Automated diagnosis using deep learning on chest X-rays (CXRs) offers a promising approach.

Purpose of the Study:

• To investigate the efficacy of deep learning models, including Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), for automated cardiomegaly diagnosis from CXRs.
• To develop and evaluate a novel ensemble model for enhanced diagnostic performance.

Main Methods:

• A large, diverse dataset of posteroanterior (PA) CXR images was compiled from multiple sources (PadChest, NIH CXR, VinDr-CXR, CheXpert).
• Various pre-trained CNN architectures (VGG16, ResNet50, InceptionV3, DenseNet121, DenseNet201, AlexNet) and Vision Transformer models were trained and compared.
• A stacking-based ensemble model, the Combined Ensemble Learning Model (CELM), was introduced, integrating CNN features via a meta-classifier.

Main Results:

• The CELM demonstrated superior diagnostic performance with 92% accuracy, 99% precision, 89% recall, 0.94 F1-score, 92.0% specificity, and 0.90 AUC.
• The model showed high agreement with expert annotations, indicating its potential for reliable clinical application.
• Vision Transformers provided competitive results, suggesting their utility as complementary tools to CNNs.

Conclusions:

• The proposed CELM framework shows potential as an efficient and scalable decision-support tool for cardiomegaly screening.
• Further validation is recommended for its implementation in resource-limited settings like ICUs and EDs.
• Automated deep learning approaches can significantly aid in rapid and accurate cardiomegaly diagnosis.