Towards Automated Segmentation of Cardiac Scars through Multimodal AI

Myocardial scarring following myocardial infarction (MI) is a major determinant of adverse cardiac outcomes, including ventricular arrhythmias, heart failure, and sudden cardiac death. Accurate quantification of scar burden and spatial distribution is therefore essential for risk stratification and treatment planning. Late gadolinium enhancement cardiac magnetic resonance (LGE-CMR) provides a non-invasive reference standard for visualizing infarcted myocardium; however, clinical assessment remains largely dependent on manual segmentation, which is time-consuming, subjective, and difficult to scale. This limits its applicability in time-sensitive clinical workflows where results must be delivered rapidly. Existing automated approaches show promise but are typically developed under simplified assumptions, including homogeneous imaging protocols, fully annotated datasets, and reliance on single-modality models; conditions rarely encountered in real-world clinical practice.

This thesis investigates how multimodal artificial intelligence (AI) can enable robust and clinically applicable myocardial scar segmentation by progressively addressing these real-world constraints. First, an automated deep learning framework for multi-sequence CMR is introduced, demonstrating that integrating complementary contrasts yields more accurate and consistent segmentation of infarcted tissue than single-sequence approaches. Second, to mitigate limited annotations and underrepresentation of rare scar patterns, a clinically guided scar synthesis strategy is proposed, by incorporating anatomical and pathological priors to generate physiologically plausible training data and improve generalization.

Third, recognizing that infarct remodeling involves both structural and electrophysiological changes, a multimodal framework is developed that integrates 12-lead electrocardiogram (ECG) signals with LGE-CMR. A temporal-aware fusion mechanism explicitly models acquisition asynchrony, enabling coherent integration of electrical and anatomical information. Finally, a domain-generalized segmentation framework is introduced to address heterogeneous imaging protocols, missing modalities, and multi-center variability, ensuring stable performance under real-world deployment conditions.

Collectively, this work advances myocardial scar segmentation from a constrained image analysis task to a clinically grounded, multimodal, and robust learning paradigm, providing a foundation for scalable and clinically deployable AI-driven cardiac assessment.