Related Experiment Video
Updated: Jun 19, 2026
02:09
Multi-modal Pulmonary Imaging: Using Complementary Information from CT and Hyperpolarized 129Xe MRI to Evaluate Lung Structure-Function
Published on: April 12, 2024
Uncertainty-Aware, End-to-End Deep Learning for Functional Lung MRI Quantification Using 129Xe and 1H MRI
Joshua R Astley1,2, Helen Marshall1,2, Laurie J Smith1
1POLARIS, School of Medicine and Population Health, The University of Sheffield, 18 Claremont Crescent, S10 2TA, Sheffield, United Kingdom.
Radiology. Cardiothoracic Imaging
|June 18, 2026
Summary
This study introduces an automated deep learning method for predicting lung ventilation defects using specialized MRI scans. The approach accurately estimates ventilation defect percentage (VDP) without manual input, offering reliable clinical insights.
Area of Science:
- Radiology and Medical Imaging
- Artificial Intelligence in Medicine
- Pulmonary Medicine
Background:
- Accurate assessment of lung ventilation is crucial for diagnosing and managing pulmonary diseases.
- Current methods for quantifying ventilation defects, such as segmentation-derived VDP, often require manual intervention, which can be time-consuming and prone to variability.
- Advanced imaging techniques like hyperpolarized 129Xe MRI offer detailed functional lung information but require sophisticated analysis pipelines.
Purpose of the Study:
- To develop and validate an end-to-end deep learning pipeline for the automatic prediction of ventilation defect percentage (VDP).
- To integrate coregistered functional 129Xe MRI and structural 1H MRI data without manual segmentation.
- To incorporate uncertainty quantification into the deep learning framework to assess prediction confidence.
Main Methods:
- A retrospective study utilizing 574 paired 129Xe MRI and 1H MRI scans from healthy individuals and patients with lung diseases.
- An uncertainty-aware convolutional neural network framework employing Monte Carlo dropout for uncertainty estimation.
- Test-time augmentation was used to evaluate model robustness and simulate test-retest repeatability.
Main Results:
- The deep learning pipeline achieved a median absolute error of 1.01% for VDP prediction, comparable to manual segmentation methods (P = .70).
- The model demonstrated strong clinical classification accuracy of 91% (95% CI: 68, 94).
- The uncertainty quantification provided confidence groupings for VDP predictions.
Conclusions:
- An automated, uncertainty-aware deep learning approach can accurately predict VDP from multi-modal MRI without manual segmentation.
- The developed framework offers performance comparable to traditional segmentation-based methods.
- This approach enhances the efficiency and reliability of functional lung imaging analysis.
