Related Experiment Video
Updated: Aug 15, 2025
07:53
Author Spotlight: Advancing 3D Modeling for Enhanced Diagnosis and Treatment of Pulmonary Nodules in Early-Stage Lung Cancer
Published on: October 13, 2023
1.5K
A radiographic, deep transfer learning framework, adapted to estimate lung opacities from chest x-rays
Avantika Vardhan1,2, Alex Makhnevich1,3, Pravan Omprakash2
1Institute of Health System Science, Feinstein Institutes for Medical Research, Northwell Health, Manhasset, NY, 11030, USA.
Bioelectronic Medicine
|January 3, 2023
Summary
A deep transfer learning framework was developed to estimate lung opacity from chest radiographs (CXRs). The ResNet-50 model demonstrated superior agreement with radiologist scores, outperforming inter-radiologist consistency.
Area of Science:
- Radiology and Medical Imaging
- Artificial Intelligence in Healthcare
- Deep Learning for Medical Diagnostics
Background:
- Chest radiographs (CXRs) are crucial for detecting respiratory diseases causing lung opacities.
- Current CXR opacity estimates often lack standardization, leading to subjective and irreproducible results.
Purpose of the Study:
- To develop a robust deep transfer learning framework for estimating lung opacity degree from CXRs.
- To adapt and validate the framework using ordinal multiclass classification.
Main Methods:
- A dataset of 38,365 prospectively annotated CXRs was utilized.
- ResNet-50, VGG-16, and ChexNet architectures were evaluated with various segmentation and data balancing strategies.
- Model performance was assessed using metrics like MAE and a novel Macro-Averaged Heatmap Concordance Score (MA HCS).
Main Results:
- The ResNet-50 model combined with undersampling and no Region Of Interest (ROI) segmentation yielded optimal MAE and HCS.
- The developed model's performance metrics showed superior agreement with radiologist scores compared to inter-radiologist agreement.
- Sensitivity analysis was performed across diverse patient populations.
Conclusions:
- The proposed deep transfer learning framework effectively estimates lung opacity degree from CXRs.
- The ResNet-50 model offers a reliable and reproducible method for lung opacity assessment, potentially improving diagnostic consistency.
