Beyond segmentation: an uncertainty-aware end-to-end approach to functional lung image quantification

Motivation: Ventilation defect percent (VDP) is a widely-used and highly-sensitive biomarker for quantifying ventilation. VDP is computed from paired functional hyperpolarized xenon-129 (129Xe)-MRI and structural 1H-MRI lung segmentations. Despite advances in deep learning (DL)-based segmentation methods, time-consuming manual editing is still required. Goal(s): To predict VDP directly from 129Xe- and 1H-MRI without image segmentation. Approach: We propose an uncertainty-aware end-to-end neural network trained and tested on 574 pairs of 129Xe- and 1H-MRI. Monte-Carlo dropout was used to quantify probabilistic uncertainty. Results: End-to-end DL predictions show statistical equivalence with segmentation-based workflows. Uncertainty quantification provides important information about the trustworthiness of predictions. Impact: Direct prediction of 129Xe-MRI lung ventilation metrics using a multi-modality dual-channel convolutional neural network for end-to-end functional lung image quantification. Uncertainty quantification is utilized to improve trust and reliability in predictions and facilitate streamlined triaging in functional lung imaging workflows.