Interactive Deep Learning for Myocardial Scar Segmentation Using Cardiovascular Magnetic Resonance.
Authors
Affiliations (6)
Affiliations (6)
- Department of Cardiovascular Sciences, University of Leicester, the National Institute for Health and Care Research Leicester Biomedical Research Centre and British Heart Foundation Centre of Research Excellence, Glenfield Hospital, Leicester, UK.
- Department of Information Engineering and Mathematics, University of Siena, Siena, Italy.
- Department of Mathematics, University of Leicester, Leicester, UK.
- Department of Cardiovascular Sciences, University of Leicester, the National Institute for Health and Care Research Leicester Biomedical Research Centre and British Heart Foundation Centre of Research Excellence, Glenfield Hospital, Leicester, UK; Centre for Digital Health and Precision Medicine, University of Leicester.
- Pioneer Centre for AI, Department of Computer Science, University of Copenhagen, Copenhagen, Denmark.
- Department of Cardiovascular Sciences, University of Leicester, the National Institute for Health and Care Research Leicester Biomedical Research Centre and British Heart Foundation Centre of Research Excellence, Glenfield Hospital, Leicester, UK. Electronic address: [email protected].
Abstract
Following myocardial infarction, late gadolinium-enhancement (LGE) assessed by cardiovascular magnetic resonance (CMR) provides a reliable metric for risk stratification and therapeutic planning. However, conventional segmentation methods are time-consuming and labor-intensive, with high interobserver variability and inconsistent performance in routine clinical practice. This study sought to develop an interactive deep learning system for scar segmentation and quantification. The framework was developed and evaluated using LGE-CMR images from 348 patients with chronic myocardial infarction (244 training, 51 validation, 53 test). The model incorporates prompt-guided segmentation and leverages a vision foundation model adapted for medical imaging, integrated into a clinician-facing interface for real-time interaction, and automated quantification. Training used a composite loss function combining Dice overlap, voxel-wise cross-entropy, and Kullback-Leibler divergence against soft labels to address annotation uncertainty. Performance was evaluated on a held-out test set using expert manual annotations as the reference standard, with assessment of segmentation accuracy, repeatability, and agreement with the conventional full-width at half-maximum method (FWHM). The framework achieved expert-level segmentation performance on the test set (Dice similarity coefficient=0.74±0.10; Hausdorff distance=5.87±6.79mm) with median scar mass error of 1.28g (IQR 0.74-2.34), corresponding to 1.4% (IQR 0.81-2.47) of left ventricular mass. Repeatability analysis (n=41) demonstrated excellent agreement, with both inter- and intra-observer concordance correlation coefficients of 0.999 (compared with 0.737 and 0.952, respectively, for the conventional FWHM). Segmentation time was substantially reduced when using the interactive tool compared with the conventional workflow, averaging 65 ± 34seconds per patient. Performance and repeatability remained high across the test set with differing levels of image quality. The proposed framework for scar segmentation with a human-in-the-loop design enables fast, accurate, and highly reproducible myocardial scar quantification from LGE-CMR. This may provide more consistent performance in routine clinical workflows.