Optimized automated cardiac MR scar quantification with GAN-based data augmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Late gadolinium enhancement (LGE) cardiac MRI lacks standardization and requires time-consuming postprocessing for scar quantification.
• This limits the clinical utility of LGE-CMR for assessing myocardial scar.

Purpose of the Study:

• To develop and evaluate a cascaded deep learning pipeline for automated scar quantification in cardiac MRI.
• To test the hypothesis that synthetic data augmentation improves model accuracy and robustness.

Main Methods:

• A three-stage cascaded deep learning pipeline was proposed, involving bounding box regression and segmentation networks (nnU-Net).
• The models were trained on EMIDEC challenge data, augmented with a synthetic dataset generated via conditional GAN.
• The pipeline was designed to segment left ventricular myocardium and scar.

Main Results:

• The cascaded pipeline significantly outperformed a single nnU-Net for both myocardium and scar segmentation (DSC 0.84 vs 0.63 and 0.72 vs 0.46, respectively).
• Synthetic data augmentation improved scar segmentation DSC by 0.06.
• The final pipeline achieved a mean DSC of 0.86 for myocardium and 0.67 for scar on the challenge test set.

Conclusions:

• A cascaded deep learning pipeline augmented with synthetic data achieves myocardium and scar segmentation comparable to manual operators.
• This approach outperforms direct segmentation methods, offering a more robust and accurate solution for scar quantification.