Aleatoric uncertainty in accelerated prostate MRI reconstruction: echo-train dropout versus Gaussian noise Monte Carlo sampling

PURPOSE: Although undersampling combined with deep learning (DL)-based reconstruction shortens MRI acquisition, it increases the chance of inaccuracies, highlighting the need for quantifiable uncertainty measures. Two inference-time perturbation strategies, echo-train dropout (ET-Drop) and Gaussian noise Monte Carlo sampling (GN-MC), were compared in terms of the correlation between their variance-based uncertainty maps and absolute reconstruction error in DL-accelerated T2w prostate MRI.

METHODS: This retrospective multi-center study used a publicly available dataset with 312 k-spaces from NYU for training and a dataset with 120 k-spaces from University Medical Center Groningen for external validation. Fully sampled 3 T data were retrospectively undersampled to acceleration factors R = 3 and R = 6 and reconstructed by a vSHARP model. Per slice, five GN-MC perturbations were reconstructed by adding complex noise at 2.5σ, and five ET-Drop perturbations, created by omitting non-central echo trains. Voxel-wise aleatoric uncertainty was defined as the variance (σ 2) across these reconstructions and correlated with absolute reconstruction error over whole slices and within the prostate.

RESULTS: Both uncertainties yielded moderate slice-level correlations with absolute error. At R = 3, ET-Drop slightly outperformed GN-MC (median ρ = 0.39 vs 0.35; p < 0.001). At R = 6, the ranking reversed (0.44 vs 0.40; p < 0.001). Correlations within the prostate fell to 0.10-0.15. ET-Drop variance maps were dominated by coil sensitivities.

CONCLUSION: Both perturbation strategies yield variance-based uncertainty maps that correlate moderately with voxel-wise error. More importantly, they consistently highlighted acquisition-related fragility, supporting the role of uncertainty mapping as a useful quality-control tool in prostate MRI.