Stroke-Aware CycleGAN: Improving Low-Field MRI Image Quality for Accurate Stroke Assessment

Area of Science:

• Medical Imaging
• Artificial Intelligence

Background:

• Low-field portable MRI (pMRI) offers accessible imaging but produces low-resolution images, hindering stroke diagnosis.
• Accurate lesion quantification is crucial for effective stroke management and treatment planning.

Purpose of the Study:

• To develop a 3D deep learning model (SA-CycleGAN) to enhance low-field MRI image quality for stroke diagnosis.
• To improve the diagnostic accuracy and lesion quantification capabilities of pMRI devices.

Main Methods:

• Proposed SA-CycleGAN, a 3D deep learning model based on CycleGAN, incorporating stroke lesion priors via spatial feature transformation.
• Integrated gradient difference losses to mitigate over-smoothing in synthesized images.
• Utilized a dataset of 101 paired high-field and low-field diffusion-weighted imaging (DWI) scans.

Main Results:

• SA-CycleGAN generated images with significantly higher quality and clarity compared to original low-field DWI.
• Lesion volume quantification showed a strong correlation (R=0.852) between SA-CycleGAN generated images and high-field images, outperforming low-field images (R=0.462).
• Mean absolute difference in lesion volumes was significantly smaller for SA-CycleGAN (1.73±2.03 mL) compared to low-field images (2.53±4.24 mL).

Conclusions:

• SA-CycleGAN effectively enhances low-field MRI images, improving visual clarity and consistency with high-field scans.
• The model offers a cost-effective method for improving stroke diagnosis efficiency and accuracy in clinical settings.
• SA-CycleGAN demonstrates potential for wider adoption of pMRI in routine stroke care.