Automated cardiovascular MR myocardial scar quantification with unsupervised domain adaptation

Area of Science:

• Medical Imaging and Artificial Intelligence
• Cardiovascular Magnetic Resonance Imaging
• Machine Learning in Healthcare

Background:

• Automated quantification of myocardial scar using artificial intelligence (AI) in late gadolinium enhancement (LGE) cardiovascular magnetic resonance (CMR) images is crucial for clinical assessment.
• AI models often face performance degradation due to domain shifts, where data characteristics differ from the original training set, limiting their generalizability to local hospital data.
• Existing AI pipelines for scar quantification require adaptation to new datasets, often involving extensive manual annotation, which is time-consuming and resource-intensive.

Purpose of the Study:

• To investigate the feasibility of using unsupervised CycleGAN-based domain adaptation to enable AI models trained on public LGE CMR data to perform scar quantification on local hospital data.
• To evaluate the performance of an AI scar quantification pipeline after domain adaptation on an external test set of 44 patients with clinically assessed ischemic scar.
• To demonstrate that AI models trained on public datasets can be effectively deployed on institution-specific clinical data with varying acquisition settings, minimizing manual labor.

Main Methods:

• CycleGAN models were employed to translate local hospital LGE CMR data to match the appearance of a public LGE CMR dataset, facilitating domain adaptation.
• A pre-developed AI scar quantification pipeline, including myocardium segmentation, scar segmentation, and scar burden computation, was applied to the domain-adapted local data.
• The performance of the adapted AI pipeline was assessed using Dice similarity coefficients for segmentation accuracy and Bland-Altman analysis for scar burden bias on an external test set.

Main Results:

• The AI scar quantification pipeline achieved mean Dice similarity coefficients of 0.76 ± 0.05 for myocardium and 0.75 ± 0.32 for scar, comparable to previous reports.
• Performance for scar segmentation in scans with pathological findings was 0.41 ± 0.12, indicating robust quantification even in complex cases.
• Bland-Altman analysis revealed a mean bias of -0.62% in scar burden percentage with limits of agreement from -8.4% to 7.17%, demonstrating clinical applicability.

Conclusions:

• Unsupervised CycleGAN-based domain adaptation is a feasible approach for deploying AI models trained on public LGE CMR data for scar quantification on local clinical data.
• This method successfully bridges the domain gap, allowing AI tools to perform reliably on data with different acquisition characteristics without requiring additional manual annotation.
• The study highlights the potential of domain adaptation techniques to enhance the clinical utility and widespread adoption of AI-driven cardiovascular image analysis.