Exact Diffusion Reconstruction Boosts DBT by 248x, Calibrates Uncertainty

Imade Bouftini presents a novel approach to limited-angle digital breast tomosynthesis (DBT) reconstruction that enforces exact data consistency using conditional diffusion priors. The key innovation replaces the typical proximal update with exact Euclidean projection onto the data-consistent set, computed via a dual system with a one-time Gram matrix factorization, costing only 4.5 ms per step (248x speedup) and driving data residual to double-precision floor (2.4e-13). The work proves this projection is the limit of the proximal step and shows that exactly consistent sample ensembles have variance only in the unmeasured subspace, enabling a calibrated uncertainty map.

Experiments on patient-derived breast phantoms demonstrate improved fidelity with no depth-resolution cost. Isotonic recalibration brings ensemble spread to calibrated error scale (ECE 0.029 to 0.008; standardized error 4.7 to 0.96), outperforming the pure prior. The study also fixes a 20.3% adjoint mismatch in a deployed projector via a materialized operator of record. This marks the first data-consistent, uncertainty-calibrated learned reconstruction for limited-angle DBT, with the solver naturally relaxing to discrepancy-ball and maximum-a-posteriori modes for noisy measurements.