Multiview Cardiac MRI Segmentation Using a Channel-Enhanced Feature Pyramid Network

Abstract

Precise heart chamber segmentation from cardiac magnetic resonance imaging (CMR) is vital for diagnosing and monitoring cardiovascular diseases (CVDs), the leading cause of death worldwide. Current deep learning models, like as U-Net and Swin Transformer, perform well in CMR segmentation, but they require millions of parameters, which increases computing costs and limits real-time efficiency. A multiview compact Channel-wise Feature Pyramid Network (CFPNet) addresses these CMR segmentation challenges effectively and precisely. Encoder-decoder architecture with enhanced multi-scale feature fusion collects local and global contextual information in CFPNet. The Channel-wise Feature Pyramid (CFP) module, inspired by MultiResNet, uses parallel FP channels with different dilation rates and skips connections. This architecture improves multi-level feature extraction while reducing trainable parameters. The network has 835,050 parameters, fewer than U-Net (25M) and Swin Transformer (26M), but it is more efficient. After training and testing on ACDC-17, M&M, and M&M II, the proposed network achieved high Dice scores of 0.94, 0.893, and 0.937 for short-axis views and 0.93 for long-axis views. Validation under various imaging circumstances and novel CVD cases shows the model's robustness and generalizability. These results demonstrate the proposed method's viability for real-time, resource-efficient CMR segmentation in clinical practice.