Cardiac function in a large animal model of myocardial infarction at 7 T: deep learning based automatic segmentation...

Area of Science:

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

Background:

• Cardiac magnetic resonance (CMR) imaging is crucial for non-invasive cardiac function quantification.
• Accurate myocardial segmentation is essential for CMR analysis but current clinical software struggles with preclinical and ultrahigh field strength data (e.g., 7 Tesla).
• Preclinical studies often require specialized segmentation methods due to differences in image acquisition and characteristics compared to human clinical data.

Purpose of the Study:

• To adapt and re-train an established deep learning (DL) model for myocardial segmentation.
• To improve the speed and reproducibility of cardiac function analysis using preclinical 7 Tesla CMR data.
• To evaluate the performance of the re-trained DL model against manual segmentation in a preclinical setting.

Main Methods:

• A deep learning model, initially developed for human CMR segmentation, was re-trained using 772 manually labeled preclinical 7 Tesla CMR images from eleven pigs.
• The re-trained model was then tested on a separate set of 288 images.
• Inter- and intra-observer variability of manual segmentation was assessed for comparison.

Main Results:

• The re-trained DL model demonstrated excellent agreement with manual segmentation for key cardiac function parameters.
• Ejection fraction (EF) analysis showed a Pearson's r of 0.95, an Intraclass correlation coefficient (ICC) of 0.97, and a Coefficient of variability (CoV) of 6.6%.
• Segmentation accuracy was high, with Dice scores of 0.88 for the left ventricle and 0.84 for the myocardium.

Conclusions:

• Re-training a human-focused DL model on preclinical 7 Tesla data significantly enhances myocardial segmentation accuracy.
• The adapted DL model provides a faster and more reproducible method for cardiac function analysis in preclinical CMR studies.
• This approach overcomes limitations of existing clinical software for ultrahigh field strength preclinical imaging.