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

Joshua R Astley, Helen Marshall, Laurie J Smith, Alberto M Biancardi, Guilhem J Collier, Laura C Saunders, Jim M Wild, Bilal Tahir

Keywords: Functional imaging, lung, multi-modal, MRI, pulmonary, uncertainty-aware.

TL;DR: We develop an uncertainty-aware, dual-channel CNN-based framework to directly predict a key functional lung imaging metric, without segmentation, from multi-modal structural and functional volumetric scans.

Abstract: Functional lung imaging modalities, such as hyperpolarized gas MRI, facilitate the visualization and quantification of regional lung ventilation. The ventilation defect percentage (VDP) is a highly-sensitive biomarker for quantifying small changes in lung function, derived from spatially co-registered functional hyperpolarized Xenon-129 ($^1$$^2$$^9$Xe)-MRI and structural proton ($^1$H)-MRI. However, manual-editing associated with segmentation-based workflows represents a time-consuming obstacle to delivering functional lung MRI results to clinicians. End-to-end deep learning (DL), which predicts final outputs without intermediary steps, frequently demonstrates improved performance on computer vision tasks; however, intermediary steps can no longer be interrogated. In this work, we developed the first end-to-end, uncertainty-aware DL framework for directly predicting VDP and its associated confidence. The direct prediction of VDP can potentially provide clinicians with important clinical data faster than segmentation-based methods.

Submission Number: 30