Abstract
Convolutional Neural Networks (CNNs) are powerful tools for image reconstruction but often operate as "black boxes," making it difficult to distinguish between real features and reconstruction artifacts. This paper presents SUREMap, a novel approach that utilizes Stein’s Unbiased Risk Estimate (SURE) to generate per-pixel uncertainty heatmaps for CNN-based reconstructions. Unlike previous methods that focus solely on variance, SUREMap estimates the total Mean Squared Error (MSE)—including both bias and variance—without requiring access to ground-truth images. By integrating SURE with Approximate Message Passing (AMP) frameworks, the proposed method provides a statistically grounded way to visualize potential errors in safety-critical applications such as medical imaging.
Introduction
In computational imaging domains like Magnetic Resonance Imaging (MRI) and compressive sensing, CNNs frequently outperform traditional mathematical models. However, their propensity for "hallucinations"—generating plausible-looking but incorrect features—limits their reliability in clinical settings. A clinician must be able to determine if a detected feature is a biological reality or a reconstruction artifact.
While Mean Squared Error (MSE) is the standard metric for reconstruction quality, it typically requires a "ground truth" image, which is unavailable in practice. Existing uncertainty quantification methods, such as those based on Bayesian dropout or bootstrapping, often only capture model variance and ignore systematic bias. SUREMap addresses this gap by providing a reliable estimate of the total per-pixel MSE using only noisy measurement data.
Methodology
The core of SUREMap lies in the synergy between Stein’s Unbiased Risk Estimate (SURE) and Approximate Message Passing (AMP) solvers.
The AMP Framework
Iterative solvers like AMP and its variant, Variable Density AMP (VDAMP), decompose complex reconstruction problems into a sequence of simpler denoising tasks. A critical property of AMP is that it ensures the "effective noise" at each iteration follows a Gaussian distribution. This mathematical guarantee is essential for the application of SURE.
Stein’s Unbiased Risk Estimate (SURE)
SURE is a statistical formula that provides an unbiased estimate of the MSE of a denoiser, provided the input noise is Gaussian. By operating within the AMP framework, SUREMap can be applied to the CNN-based denoisers used in the reconstruction loop.
Monte-Carlo Divergence Estimation
Calculating SURE requires the divergence of the denoiser function. Since CNNs lack a closed-form derivative, the authors employ a Monte-Carlo approximation. This involves perturbing the network input with infinitesimal noise and observing the output change to estimate the divergence.
Patch-based Heatmaps
To stabilize the uncertainty estimates, SURE is computed over overlapping image patches (e.g., 48x48 pixels) and averaged. This process produces a smooth, interpretable heatmap of the predicted error across the image.
Experimental Results
The researchers evaluated SUREMap using natural images and medical MR images. They utilized the DnCNN architecture, a widely used residual learning denoiser, trained on a range of noise levels.
Visual and Quantitative Fidelity
As shown in the figure above, the SURE-generated heatmaps closely align with the "True MSE" maps generated using ground-truth data. Quantitatively, the normalized absolute difference between the SURE estimate and the actual MSE decreases as the patch size increases, confirming the statistical validity of the approach. The method proves versatile, performing well for both standard compressive sensing and complex MRI reconstructions involving colored noise in the wavelet domain.
Key Insights
- Bias + Variance Capture: Unlike sensitivity-based methods, SUREMap captures the actual expected error, including the bias introduced by the CNN's learned priors.
- Resolution vs. Accuracy: There is an inherent trade-off in heatmap generation. Smaller patches offer higher spatial resolution but higher statistical noise, while larger patches provide more accurate estimates at the cost of "blurrier" heatmaps. The authors found that 48x48 pixel patches provided the best balance.
Conclusion and Future Work
SUREMap provides a robust framework for quantifying uncertainty in CNN-based image reconstruction. While the current implementation faces a trade-off between heatmap resolution and computational overhead due to Monte-Carlo iterations, it represents a significant step toward making deep learning models more transparent in high-stakes environments. Future research may explore using these heatmaps for automated artifact removal or guiding adaptive sampling strategies in MRI to prioritize data collection in high-uncertainty regions.
Summary updated