Related Experiment Video
Updated: Jun 19, 2026
09:16
Supervised Machine Learning for Semi-Quantification of Extracellular DNA in Glomerulonephritis
Published on: June 18, 2020
6.9K
Automatic Quantification of COVID-19 Pulmonary Edema by Self-supervised Contrastive Learning
Zhaohui Liang1, Zhiyun Xue1, Sivaramakrishnan Rajaraman1
1Computational Health Research Branch, National Library of Medicine, National Institutes of Health, Bethesda, MD, USA.
Medical Image Learning with Limited and Noisy Data : Second International Workshop, Milland 2023, Held in Conjunction with MICCAI 2023, Vancouver, BC, Canada, October 8, 2023, Proceedings. Milland (Workshop) : (2Nd : 2023 : Vancouver, B
|February 28, 2024
Summary
A new self-supervised machine learning model accurately rates pulmonary edema severity on chest X-rays using the mRALE score. This AI approach shows superior performance compared to traditional methods, aiding in COVID-19 pneumonia assessment.
Area of Science:
- Artificial Intelligence
- Medical Imaging
- Machine Learning
Background:
- Pulmonary edema assessment on chest X-rays (CXR) is crucial for diagnosing conditions like COVID-19 pneumonia.
- Automated scoring systems can improve efficiency and consistency in evaluating lung edema severity.
Purpose of the Study:
- To develop and evaluate a self-supervised machine learning method for automated pulmonary edema severity rating using the modified radiographic assessment of lung edema (mRALE) scoring system on frontal CXRs.
- To compare the performance of the proposed model against non-pretrained and from-scratch trained comparators.
Main Methods:
- A self-supervised contrastive learning approach utilizing a Siamese network (SimSiam) architecture with a pre-trained ResNet-50 backbone was employed.
- The model generated 2048-dimension embeddings for a downstream deep neural network to predict mRALE scores.
- Performance was evaluated using 5-fold cross-validation on 2,599 frontal CXRs and external validation.
Main Results:
- The proposed model achieved superior performance with a mean absolute error (MAE) of 5.05, mean squared error (MSE) of 66.67, and Spearman's correlation coefficient (Spearman ρ) of 0.77 compared to comparators (P<0.01).
- External validation demonstrated a prediction probability concordance of 0.811 and quadratic weighted kappa of 0.739.
- Comparator models showed no statistically significant difference in MSE and Spearman ρ (P>0.05).
Conclusions:
- Self-supervised contrastive learning is an effective strategy for automated mRALE scoring.
- The developed method offers a novel approach to enhance machine learning performance in quantitative medical image analysis.
- This technique minimizes the need for extensive expert knowledge in learning patterns from medical images.
