Deep learning-based left ventricular segmentation demonstrates improved performance on respiratory motion-resolved...

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medicine
• Medical Image Analysis

Background:

• Deep learning (DL) based segmentation is increasingly used for cardiac magnetic resonance (CMR) analysis, particularly for left ventricular (LV) volume determination.
• Whole-heart, free-breathing, self-navigated CMR offers high-resolution imaging for cardiac anatomy assessment.
• Combining advanced CMR acquisition with DL segmentation can enhance clinical workflow efficiency.

Purpose of the Study:

• To compare the performance of a DL automatic LV segmentation network, trained on CT images, across two whole-heart CMR reconstruction methods: motion-corrected (Mcorr) and motion-resolved (Mres).
• To evaluate if DL segmentation generalizes better to Mres images, hypothesized to have superior image quality over Mcorr images.
• To assess the accuracy of DL-based LV volume quantification against manual expert tracings.

Main Methods:

• Retrospective analysis of 15 patients undergoing 1.5T CMR using a 3D radial phyllotaxis bSSFP sequence.
• Comparison of DL-based automatic LV segmentation with manual expert tracings using Absolute Volume Difference (AVD) and 3D Dice Similarity Coefficient.
• Assessment of LV myocardium-blood pool interface sharpness and statistical analysis using paired Student's t-test.

Main Results:

• Lower Absolute Volume Difference (AVD) for Mres (7.73 ± 6.54 ml) compared to Mcorr (20.0 ± 22.4 ml) reconstructions (p=0.03).
• Higher 3D Dice Similarity Coefficient for Mres (0.90 ± 0.02) versus Mcorr (0.87 ± 0.03) images (p=0.02).
• Significantly greater sharpness observed in Mres images (0.15 ± 0.05) than Mcorr images (0.12 ± 0.04) (p=0.014).

Conclusions:

• The DL-based 3D automatic LV segmentation network generalized better to Mres images than Mcorr images for LV volume quantification.
• The findings support the hypothesis that Mres reconstructions yield more accurate DL-based LV segmentation.
• This study highlights the potential of Mres reconstruction to improve the reliability of automated LV volume analysis in whole-heart CMR.