CapNet: An Automatic Attention-Based with Mixer Model for Cardiovascular Magnetic Resonance Image Segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Research

Background:

• Deep neural networks excel in medical image segmentation, particularly for cardiac MRI.
• Transformer models offer long-range dependency learning but suffer from high computational costs and parameter counts.
• Existing methods often require pre-training on large datasets, increasing memory and resource demands.

Purpose of the Study:

• To introduce CapNet, a novel, lightweight, and efficient convolutional neural network for cardiac MRI segmentation.
• To address the challenges of varying cardiac shapes and sizes during different phases.
• To propose a new loss function, the Tversky Shape Power Distance, for improved segmentation accuracy.

Main Methods:

• Developed CapNet, a convolutional model with mixing modules, trainable from scratch with minimal parameters.
• Incorporated attention modules for pooling, spatial, and channel information to handle shape variations.
• Utilized the Tversky Shape Power Distance loss function to emphasize shape dissimilarity.

Main Results:

• Achieved high Dice Similarity Coefficients (DSC) on public datasets: up to 96.82% for endocardium/epicardium segmentation.
• Demonstrated strong performance in multiclass segmentation, with average DSC of 93.05% on ACDC and high scores across MS-CMR sequences.
• CapNet showed statistically significant improvements over transformer-based and CNN-based methods despite having fewer parameters.

Conclusions:

• CapNet offers a lightweight yet highly efficient solution for cardiac MRI segmentation.
• The proposed attention and loss function modules effectively handle cardiac shape variability.
• CapNet provides competitive and statistically significant results compared to state-of-the-art methods, reducing computational burden.