Realistic noise synthesis reduces bias and improves tissue microstructure estimation with supervised machine learning
Researchers have developed a Realistic Noise Synthesis (RNS) framework to improve the accuracy of microstructure estimation in diffusion MRI. This method addresses a bias introduced when machine learning models trained on simulated data encounter different noise characteristics in real-world MRI scans. By incorporating Rician expectation and effective post-processing noise variance into simulated training data, RNS significantly reduces parameter bias, especially in low signal-to-noise ratio (SNR) conditions. AI
IMPACT Enhances the precision of AI models in medical imaging, particularly for low-SNR diffusion MRI data.