Deep Learning Single-Frame and Multiframe Super-Resolution for Cardiac MRI.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Cardiac MRI acquisition is time-consuming, limiting spatial resolution.
• Faster MRI scans with smaller matrices reduce image detail.
• Deep learning (DL) offers potential for faster acquisition and enhanced spatial detail via super-resolution.

Purpose of the Study:

• To assess the feasibility of using DL for super-resolution in cardiac MRI.
• To compare DL performance against conventional image upscaling techniques.

Main Methods:

• Convolutional neural networks (CNNs) were trained for super-resolution using cardiac MRI data.
• CNNs were compared to bicubic interpolation and zero padding using Structural Similarity Index (SSIM).
• Clinical performance was evaluated by measuring left ventricular volumes.

Main Results:

• CNNs significantly outperformed traditional methods (bicubic interpolation, zero padding) in upsampling MRI images (P < .001).
• Super-resolved images yielded left ventricular volumes comparable to full-resolution images (P > .05).
• DL enhanced anatomic detail in cardiac MRI, even in imaging planes not used for training.

Conclusions:

• Deep learning surpasses conventional upscaling methods for cardiac MRI.
• DL effectively recovers high-frequency spatial information, improving image quality.
• The DL strategy shows promise for enhancing MRI quality across different imaging planes.