Development and performance evaluation of fully automated deep learning-based models for myocardial segmentation on...

Area of Science:

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

Background:

• Accurate segmentation of the left ventricular (LV) myocardium is crucial for quantitative analysis of cardiac magnetic resonance imaging (CMR) T1 maps.
• Manual segmentation is time-consuming and subject to inter-observer variability.
• Deep learning offers a potential solution for automated and accurate myocardial segmentation.

Purpose of the Study:

• To develop and optimize a deep learning-based model for segmenting the LV myocardium on native T1 maps.
• To evaluate the model's performance in both long-axis and short-axis cardiac MRI orientations.
• To compare the model's segmentation accuracy against manual segmentation by human raters.

Main Methods:

• A U-Net architecture was employed for deep learning model development.
• Models were trained on native myocardial T1 maps from 125 participants (50 healthy volunteers, 75 patients).
• Systematic optimization involved different training metrics (DSC, IOU), activation functions (ReLU, LeakyReLU), and epochs; 35 epochs with DSC and ReLU yielded optimal results.

Main Results:

• The optimized deep learning model achieved high segmentation performance (mean DSC 0.88 ± 0.07) with a mean T1 error of 10.6 ± 17.9 ms.
• The model's limits of agreement (-35.5 to +36.1 ms) were superior to those between two human raters (-34.7 to +59.1 ms).
• Segmentation accuracy was comparable for both long-axis (mean DSC 0.89 ± 0.03) and short-axis (mean DSC 0.88 ± 0.08) views.

Conclusions:

• Fully automated segmentation of the LV myocardium on native T1 maps is feasible using deep learning.
• The developed model demonstrates high accuracy and consistency, surpassing human inter-observer agreement.
• This automated approach enables efficient and reliable quantitative analysis of myocardial T1 maps in clinical settings.