Convolutional recurrent U-net for cardiac cine MRI reconstruction via effective spatio-temporal feature exploitation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Cardiac Cine Magnetic Resonance Imaging (MRI) offers dynamic heart visualization but is limited by slow acquisition times, necessitating breath-holds.
• Accelerated imaging techniques can speed up scans but may compromise image resolution.
• Deep learning models show promise for reconstructing high-quality images from undersampled data, but often fail to fully exploit spatio-temporal features in cine MRI.

Purpose of the Study:

• To enhance cine MRI reconstruction by integrating convolutional recurrent operations with a U-Net architecture.
• To improve the exploitation of inherent spatio-temporal features for superior reconstruction performance.

Main Methods:

• Developed CRUNet-MR, a deep learning model combining convolutional recurrent operations and U-Net architecture for enhanced spatio-temporal feature extraction.
• Incorporated dilated convolutions to broaden the spatial receptive field and explored optimal dilation factors.
• Trained and validated the model on the CMRxRecon2023 dataset with varying acceleration factors (4-24) and tested on an in-house dataset.

Main Results:

• CRUNet-MR demonstrated statistically significant improvements over benchmark models, particularly in reconstructing dynamic cardiac regions.
• Achieved high reconstruction quality with average SSIM scores of 0.986 at 4x and 0.971 at 8x acceleration.
• Ablation studies confirmed the effectiveness of the model's design choices, and validation on an in-house dataset showed good generalization and adaptability.

Conclusions:

• CRUNet-MR effectively reconstructs high-quality cine MRI images, especially in dynamic cardiac areas, by leveraging spatio-temporal features.
• The model's performance supports higher acceleration factors, paving the way for faster and more patient-friendly cardiac imaging.