Cardiac MRI segmentation using shifted-window multilayer perceptron mixer networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Accurate segmentation of cardiac structures in magnetic resonance imaging (MRI) is crucial for diagnosing and managing cardiovascular diseases.
• Current deep learning models face challenges in achieving precise and efficient segmentation of the left ventricle, right ventricle, and myocardium.

Purpose of the Study:

• To develop and evaluate a novel deep-learning segmentation algorithm for cardiac MRI.
• To improve the accuracy and efficiency of segmenting the left ventricle, right ventricle, and myocardium (Myo).

Main Methods:

• A shifted window multilayer perceptron (Swin-MLP) mixer network was proposed, utilizing a 3D U-shaped encoder-decoder architecture.
• The network was trained and evaluated on public cardiac MRI data from 100 individuals.
• Performance was quantitatively assessed using Dice score coefficient, precision, sensitivity, Hausdorff distance (HD), mean surface distance (MSD), and residual mean square distance (RMSD).
• The proposed network was benchmarked against Dynamic UNet and Swin-UNetr algorithms.

Main Results:

• The Swin-MLP network achieved high volume similarity metrics: Dice = 0.952 ± 0.017, precision = 0.948 ± 0.016, sensitivity = 0.956 ± 0.022.
• Average surface similarity metrics were HD = 1.521 ± 0.121 mm, MSD = 0.266 ± 0.075 mm, and RMSD = 0.668 ± 0.288 mm.
• The proposed network demonstrated statistically significant improvements (p < 0.05) over Dynamic UNet and Swin-UNetr for most volumetric and surface metrics.

Conclusions:

• The Swin-MLP mixer network offers superior cardiac MRI segmentation accuracy compared to current leading methods.
• This robust algorithm has the potential to optimize clinical workflows in cardiovascular imaging.
• The findings highlight the efficacy of Swin-MLP for precise cardiac structure delineation.