Weakly-Supervised Deep learning for Left Ventricle Fibrosis Segmentation in Late Gadolinium Enhanced Cardiac MRI using Image Level Labels.

Roel Klein1, Florence van Lieshout2, Maarten Kolk3, Kylian van Geijtenbeek3, Fleur Tjong3, Romy Vos3, Samuel RuipĂ©rez-Campillo4, Ruibin Feng4, Brototo Deb4, Prasanth Ganesan4, Reinoud Knops3, Ivana Isgum5, Sanjiv Narayan4, Erik Bekkers6, Bob de Vos5
1Department of Cardiology, Amsterdam Medical Centers (location AMC), University of Amsterdam, 2Department of Cardiology, Amsterdam University Medical Centers, University of Amsterdam, 3Department of Cardiology, Amsterdam University Medical Centers (location AMC), University of Amsterdam, 4Department of Medicine, Stanford University, 5Department of Biomedical Engineering, Amsterdam Universitair Medische Centra, University of Amsterdam, 6Faculty of Science, University of Amsterdam


Abstract

Background:

Automated segmentation and quantification of myocardial fibrosis in LGE cardiac MRI (CMR) has the potential to improve efficiency and precision of diagnosis and treatment of cardiomyopathies. However, state-of-the-art Deep Learning approaches require manual pixel-level annotations, which are cumbersome to obtain, and ambiguity in interpretation results in high interobserver variability. We hypothesize that a weakly-supervised fibrosis segmentation method may be a better alternative for detection of fibrosis, allowing for expedited and scalable dataset curation and potentially greater generalizability.

Methods:

Short-axis CMRs with late gadolinium enhancement (LGE) were retrospectively obtained from 470 patients with ischemic and non-ischemic cardiomyopathies. Each LGE CMR had on average 10 image slices. For image-level detection of fibrosis a dilated residual network (DRN) was trained on 430 CMRIs with image-level fibrosis annotations, validated and tested using 20 CMRs each with pixel-precise delineation of fibrotic lesions. Lesion delineations were obtained from the network using Class Activation Maps (CAMs). To guide fibrosis predictions, a U-Net was tasked with delineation of the LV myocardium using 75 CMRs with 70/10/20 training/validation/test-split (Dice score 0.821). Subsequently, the CAMs are restricted to the U-net myocardium predictions. The model is evaluated for image-level classification and pixel-level fibrosis segmentation.

Results:

Automatic image-level fibrosis classification resulted in an Area Under the Receiver Operator Curve of 0.95 and an accuracy of 0.86. Our method, trained with only image-level labels, reached a reliable pixel-level segmentation performance with a Dice score of 0.54.

Conclusion:

Our Deep Learning models show promising results for both image-level fibrosis classification and fibrosis segmentation in LGE CMR, despite only being trained for image-level predictions. Both can provide invaluable features for subsequent automated diagnosis and risk prediction.