Weakly-Supervised Deep Learning for Left Ventricle Fibrosis Segmentation in Cardiac MRI Using Image-Level Labels

• Main Authors: Klein, Roel C, van Lieshout, Florence E, Kolk, Maarten Z, van Geijtenbeek, Kylian, Vos, Romy, Ruiperez-Campillo, Samuel, Feng, Ruibin, Deb, Brototo, Ganesan, Prasanth, Knops, Reinoud, Isgum, Ivana, Narayan, Sanjiv, Bekkers, Erik, de Vos, Bob, Tjong, Fleur V
• Format: Conference Proceeding
• Language: English
• Subjects: Annotations; Deep learning; Image segmentation; Magnetic resonance imaging; Manuals; Myocardium; Training
• ISSN: 2325-887X
• DOI: 10.22489/CinC.2022.197

Automated segmentation of myocardial fibrosis in late gadolinium enhancement (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. Using weaker labels can greatly reduce manual annotation time and expedite dataset curation, which is why we propose fibrosis segmentation methods using either slice-level or stack-level fibrosis labels. 5759 short-axis LGE CMR image slices were retrospectively obtained from 482 patients. U-Nets with slice-level and stack-level supervision are trained with 446 weakly-labeled patients by making use of a myocardium segmentation U-Net and fibrosis classification Dilated Residual Networks (DRN). For comparison, a U-Net is trained with pixel-level supervision using a training set of 81 patients. On the proprietary test set of 24 patients, pixel-level, slice-level and stack-level supervision reach Dice scores of 0.74, 0.70 and 0.70, while on the external Emidec dataset of 100 patients Dice scores of 0.55, 0.61 and 0.52 were obtained. Results indicate that using larger weakly-annotated datasets can approach the performance of methods using pixel-level annotated datasets and potentially improve generalization to external datasets.