Uncertainty-Aware, End-to-End Deep Learning for Functional Lung MRI Quantification Using 129Xe and 1H MRI

Purpose

To develop an automatic end-to-end deep learning pipeline for predicting the ventilation defect percentage (VDP) from coregistered functional hyperpolarized xenon 129 (129Xe) MRI and structural proton (1H) MRI scans without manual intervention.

Materials and Methods

In this retrospective study (2015–2024), 129Xe MRI and 1H MRI scans from healthy participants and patients with a range of pulmonary diseases were used to predict VDP and its associated prediction confidence via an uncertainty-aware convolutional neural network framework. Monte Carlo dropout was used to quantify model uncertainty. Model robustness was assessed using test-time augmentation to simulate test-retest repeatability. The proposed approach was evaluated on a stratified testing set via the median absolute error.

Results

The dataset comprised 574 paired 129Xe MRI and 1H MRI scans from 47 healthy participants (mean ± SD age, 28.3 years ± 17.3; 28 female participants) and 527 patients with a range of pulmonary pathologies (mean ± SD age, 44.9 years ± 21.9; 295 female patients). The proposed framework produced a median absolute error of 1.01% (IQR, 0.49–2.47) VDP compared with manually corrected, segmentation-derived VDPs; no evidence of difference was found (P = .70). Twenty Monte Carlo dropout iterations were completed, producing VDP prediction distributions that were subsequently clustered into confidence groupings. The proposed approach demonstrated clinical classification accuracy of 91% (95% CI: 68, 94; 32 of 35).

Conclusion

An uncertainty-aware, end-to-end deep learning approach enabled accurate prediction of VDP without manual segmentation, with performance comparable to segmentation-based methods and quantification of prediction uncertainty.