nnU-Net and MedSAM segment LV myocardium on cardiac MRE without extra MRI, reaching inter-reader accuracy

Deep learning models nnU-Net and MedSAM directly segmented the left ventricular myocardium on native cardiac MRE data, achieving Dice scores of 0.82—matching or exceeding the inter-reader agreement of 0.79—paving the way for a self-contained MRE workflow.

• nnU-Net trained on fully averaged normalized magnitude images and a semi-automated fine-tuned MedSAM both reached Dice scores of 0.82, matching inter-reader agreement (0.79).
• Using magnitude-plus-phase or real-plus-imaginary inputs substantially reduced Dice to 0.65 and 0.60, underscoring the importance of input representation.
• Fully automated MedSAM variants (autoprompt, box regression) underperformed with Dice 0.68–0.71.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.