Related Experiment Video
Updated: Dec 6, 2025
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
2.2K
A Systematic Search over Deep Convolutional Neural Network Architectures for Screening Chest Radiographs
Summary
Single deep convolutional neural network (CNN) models show promise for automating chest radiograph screening, matching ensemble performance. Xception and ResNet-18 architectures accurately identify multiple thoracic conditions, aiding healthcare in resource-limited settings.
Area of Science:
- Medical Imaging Analysis
- Artificial Intelligence in Healthcare
Background:
- Chest radiographs are crucial for diagnosing pulmonary and cardiothoracic conditions.
- The need for on-premise radiologists limits access in low-resource settings.
- Machine learning automation is being explored to overcome these limitations.
Purpose of the Study:
- To identify single deep convolutional neural network (CNN) architectures that match ensemble performance for chest radiograph screening.
- To evaluate the effectiveness of specific CNN models in detecting multiple thoracic pathologies.
Main Methods:
- Systematic search and experimentation with various standard CNN architectures.
- Conducted over 63 experiments using a 11.3 FP32 TensorTFLOPS compute system.
- Assessed model reliability using saliency maps (RISE method) and radiologist annotations.
Main Results:
- Identified Xception and ResNet-18 architectures as consistent performers.
- Achieved an average Area Under the Curve (AUC) of 0.87 across nine pathologies.
- Demonstrated that single CNN models can perform comparably to ensembles.
Conclusions:
- Single CNN models, specifically Xception and ResNet-18, are reliable for automated chest radiograph screening.
- These models show potential for improving diagnostic access in resource-limited areas.
- Limitations in the CheXpert dataset, such as class imbalance, need consideration for future research.
