Deep learning-based visualization of perfusion defects to support MRI-based perfusion scoring of the lung in cystic fibrosis (CF-MRXAI).
Project Management: Urs Eisenmann
Project partner: Translational Lung Research Center Heidelberg (TLRC), Department of Diagnostic and Interventional Radiology, Heidelberg University Hospital
Sponsor: German Center for Lung Research (DZL)
Abstract: For a systematic assessment of disease-relevant MRI findings in cystic fibrosis (CF), a scoring system was developed in 2012 by the junior group "Structural and Functional Airway Imaging" at the Translational Lung Research Center (TLRC), which accurately detects and semi-quantifies structural and functional damage in lobes. Assessment procedures include 4D perfusion sequencing with contrast injection. It allows visualization of lung perfusion failures reflecting large and small airway obstruction. Research shows that visual CF scorings by a human observer are associated with intra- and inter-reader variability that should not be underestimated. Computer-based quantitative evaluation of 4D perfusion MRI has not been satisfactorily resolved. In our own preliminary work, a first Deep Learning-based prototype for classification of perfusion defects in CF lungs was developed. The results are promising but not yet sufficient for automated perfusion scoring.
In the CF-MRXAI project, radiologists will be supported by Explainable Artificial Intelligence (XAI) methods in scoring to reduce intra- and inter-reader variability. The basic idea is to present radiologists with visualizations generated by Explainable Artificial Intelligence (XAI) alongside the perfusion sequence to allow a more objective assessment of the perfusion score. Here, reference visualizations of similar cases with comparable, slightly less pronounced or slightly more pronounced perfusion deficits are presented next to the current case to be scored. After development of an intuitive software tool, it will be evaluated by radiology staff to draw conclusions about inter- and intra-reader variability. If the tool proves useful for objective perfusion scoring, extension to other subscores important for CF diagnosis will be sought.
Duration: 03/2022 - 02/2024
Collaborators of the Institute: Friedemann Ringwald, Anna Martynova, Julian Mierisch