Cardiac function assessment with deep-learning-based automatic segmentation of free-running...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Free-running (FR) cardiac MRI offers dynamic 5D imaging but faces challenges in clinical integration due to data volume and analysis complexity.
• Current segmentation methods are inadequate for the unique spatial-temporal data of FR MRI.

Purpose of the Study:

• Develop and validate a deep learning (DL) framework for segmenting isotropic 3D+cardiac cycle FR cardiac MRI.
• Enable accurate, rapid, and clinically meaningful anatomical and functional analysis of FR cardiac MRI data.

Main Methods:

• Reconstructed 5D FR datasets from bSSFP and GRE acquisitions, retaining the end-expiratory phase for 4D datasets.
• Utilized automatic propagation of manual segmentations to train a 3D nnU-Net model for segmenting ventricular blood pools and myocardium.
• Validated using geometric, clinical, and physiological consistency metrics, including data augmentation with all cardiac phases.

Main Results:

• The DL method achieved automatic segmentation in under a minute with high geometric accuracy (DSC: 0.86-0.94) and low volume differences.
• Clinical metrics showed excellent agreement for left ventricle (LV) analysis (ICC > 0.98) and reliable RV analysis.
• Training on all cardiac phases improved temporal coherence, reducing myocardial volume mismatch from 4.0% to 2.6%.

Conclusions:

• Validated a DL-based method for fast, accurate segmentation of whole-heart FR 4D cardiac MRI.
• Demonstrated robust performance across diverse protocols, supporting integration into clinical and research workflows.
• Overcame a key barrier to the adoption of free-running cardiac MRI.