Related Experiment Video
Updated: Nov 17, 2025
08:05
Lung CT Segmentation to Identify Consolidations and Ground Glass Areas for Quantitative Assesment of SARS-CoV Pneumonia
Published on: December 19, 2020
14.5K
Deep learning models for COVID-19 chest x-ray classification: Preventing shortcut learning using feature
Caleb Robinson1, Anusua Trivedi1, Marian Blazes2
1Microsoft AI for Good.
Medrxiv : the Preprint Server for Health Sciences
|February 17, 2021
Summary
Feature disentanglement improves deep learning models for COVID-19 chest X-ray analysis. This method enhances generalization by focusing on pulmonary features, outperforming other bias reduction techniques.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- COVID-19 pandemic necessitated AI tools for chest radiograph (CXR) analysis.
- Small COVID-19+ CXR datasets and multi-source pooling lead to deep learning model overfitting (shortcut learning).
- Models learn dataset-specific biases instead of true pulmonary features.
Approach:
- Proposed feature disentanglement to train deep learning models.
- Forced models to identify pulmonary features.
- Penalized models for learning dataset-discriminating features.
Key Points:
- Feature disentanglement improved model generalization on unseen CXR data.
- Achieved up to 0.13 AUC improvement on held-out data.
- Outperformed non-lung masking and histogram equalization for bias reduction.
Conclusions:
- Feature disentanglement is an effective strategy for mitigating shortcut learning in CXR analysis.
- This approach enhances the reliability of AI models for COVID-19 diagnosis and management.
- Addresses critical data limitations in medical AI research.
