Automated Left Atrial Scar Segmentation Using a Cascaded Convolutional Neural Network

Ben A Orkild1, Jake Bergquist1, Eugene Kwan1, Eric N Paccione1, Rui Jin1, Richard Stewart Shelton2, Erik Bieging1, Lowell Chang3, Gernot Plank4, Karli Gillette5, Rob MacLeod1, Ed DiBella1, Shireen Elhabian6, Ravi Ranjan1
1University of Utah, 2Salt Lake City VA Health Care System, 3VA Salt Lake City, 4Medical University of Graz, 5Gottfried Schatz Research Center - Medical University of Graz, 6Scientific Computing and Imaging Institute, University of Utah


Abstract

Scar segmentation from late gadolinium enhanced (LGE) MRI is an essential task for noninvasive assessment of lesion formations following atrial fibrillation ablation. Several automated algorithms have been proposed, but they show poor agreement with human-segmented scars. In this study, we propose an end-to-end automated nnU-net machine learning-based pipeline to segment left atrial (LA) scar following atrial fibrillation catheter ablation. We trained and tested our convolutional neural networks (CNNs) on 49 LGE MRI scans obtained post-atrial ablation. Scar segmentations in the LGE MRIs were generated from thresholds set by 4 separate human observers, guided by ablation sites recorded during the ablation procedure. Our cascaded CNN pipeline achieved an average DICE score of 0.77 +\- 0.10 and an average surface distance of 0.36 +\- 0.17 mm on the test set (N=10). Furthermore, our performance assessment highlights the shortcomings of the DICE coefficient for LA scar segmentation.