CycleGAN models show consistent brain MRI synthesis across datasets supporting downstream tissue characterization in...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Medicine
• Neurological Disease Research

Background:

• Quantitative analysis of brain MRI is crucial for neurological diseases like multiple sclerosis (MS).
• Complete MRI datasets are often clinically unavailable, hindering analysis.
• Deep learning methods offer potential solutions for generating complete MRI datasets.

Purpose of the Study:

• To compare CycleGAN and Pix2Pix deep learning models for synthesizing brain MRI.
• To evaluate the utility of synthesized T1-weighted MRI for quantitative analysis in MS.
• To assess image quality and similarity to source data.

Main Methods:

• Utilized T1-weighted and T2-weighted brain MRI from healthy (HCP, PPMI) and MS cohorts.
• Applied bidirectional image synthesis using CycleGAN (with/without spectral normalization) and Pix2Pix.
• Performed utility testing on synthesized T1-weighted MRI, including lesion detection and volumetry.

Main Results:

• Pix2Pix generally outperformed CycleGAN on streamlined datasets, showing higher peak signal-to-noise ratios and structural similarity.
• CycleGAN performance varied with spectral normalization and dataset, improving in PPMI but not significantly in HCP or MS.
• Synthesized images demonstrated high similarity to source data in utility tests, though Pix2Pix T1 images showed more heterogeneous lesion texture.

Conclusions:

• CycleGAN without spectral normalization is feasible for synthesizing clinical brain MRI.
• Synthesized T1-weighted images are suitable for quantitative analysis in MS.
• Deep learning image synthesis shows promise for overcoming data limitations in neurological research.