Extending UQ Tools Across Self-Supervised Imaging Methods

Systematic evaluation of SURE, Tweedie, and bootstrap UQ across four SSL methods

4
UQ Tools
4
SSL Methods
3
Operator Types
6
Noise Levels

SURE Error Estimation Accuracy

Relative error of SURE estimates vs ground-truth MSE across noise levels.

SURE Relative Error vs Noise Level

Tweedie Posterior Moments

MSE improvement of Tweedie posterior mean over pseudoinverse baseline.

Tweedie MSE Improvement (%)

Equivariant Bootstrap Coverage

2-sigma coverage across operator types and SSL methods.

Coverage by Operator Type

UQ-SSL Validity Matrix

SSL MethodSURE Rel. ErrorHO-SURE Rel. ErrorTweedie MSEBootstrap CoverageVerdict
Noise2Self0.4420.4270.4990.104Partial
EI0.4560.3910.6280.313Partial
SSDU1.1001.1240.2190.250Tweedie OK
Noisier2Noise0.5500.5320.4720.250Partial

Key Findings

SURE Works Best for Noise2Self

Lowest SURE relative error among all methods due to J-invariant masking aligning with SURE assumptions.

Tweedie Improves SSDU

Tweedie posterior mean achieves lowest MSE for SSDU, where data splitting enables better score estimation.

Bootstrap Coverage Varies

Equivariant bootstrapping coverage depends heavily on operator type, with Gaussian operators yielding best results.

No Universal UQ Tool

Each UQ tool has strengths for specific SSL methods; practitioners should select based on the validity matrix.