Methods for risk map prediction in AI-based MRI reconstruction

A method for generating pixel risk maps for diagnostic image reconstruction comprising:

• feeding into a trained encoder a short-scan image acquired from a medical imaging scan to generate latent code statistics including the mean μy and variance αy;
• selecting random samples z based on the latent code statistics, where the random samples z are sampled from a normal distribution as z˜N(μy,σy2) where N represents a normal distribution function;
• feeding the random samples into a trained decoder to generate a pool of reconstructed images;
• calculating, for each pixel across the pool of reconstructed images, pixel-wise mean and variance statistics across the pool of reconstructed images;
• computing an end-to-end Jacobian of a reconstruction network that is fed with a density-compensated short-scan image, and estimating a risk of each pixel across the pool of reconstructed images, based on a Stein Unbiased Risk Estimator (SURE).