A Deep Modality-Specific Ensemble for Improving Pneumonia Detection in Chest X-rays

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Respiratory disease diagnostics

Background:

• Pneumonia diagnosis relies heavily on chest X-rays (CXRs), but current deep learning (DL) models require improvement for clinical decision support.
• Existing DL approaches for pneumonia detection in CXRs show promise but lack sufficient accuracy for widespread clinical adoption.

Purpose of the Study:

• To enhance the performance of deep learning models for detecting pneumonia in chest X-rays.
• To develop a more accurate and reliable computer-aided detection system for pneumonia using enhanced DL techniques.

Main Methods:

• Trained a DL classifier on a large CXR dataset to create a modality-specific model.
• Integrated this model as a backbone within the RetinaNet object detection network.
• Experimented with different weight initializations (random and ImageNet-pretrained) and constructed an ensemble of top-performing models.

Main Results:

• The ensemble of the top-3 RetinaNet models achieved a mean average precision (mAP) of 0.3272, significantly exceeding the state-of-the-art mAP of 0.2547.
• Ensemble methods demonstrated improved detection of pneumonia-consistent findings by reducing prediction variance.
• Optimized weight initialization strategies for classifier backbones contributed to performance gains.

Conclusions:

• The developed ensemble approach using RetinaNet with optimized DL backbones offers a substantial advancement in pneumonia detection from CXRs.
• This method shows potential to improve diagnostic accuracy and aid clinicians in timely and effective pneumonia treatment decisions.
• Further research into ensemble strategies and DL model optimization can lead to more robust AI tools for medical imaging analysis.