BreastCT-Noise leaderboard — noise-robustness re-evaluation

The Breast-CT board uses the Sidky DL-Sparse-View challenge data, which is ideal noiseless projection data by design (the RMSE floor is zero — a pure sparse-view inverse problem). This second board asks a different question: how robust is each solver to a little input noise it never trained on?

Setup (no retraining). Each solver’s noiseless best iter is re-evaluated on the same 200 held-out test cases, but with the sinograms corrupted by photon-counting (Poisson) noise at a high dose, I0 = 100 000 photons/pixel (≈1–2 % noise at the thickest ray — mild). Supervised solvers load their noiseless-trained checkpoint and skip training; per-scene / classical solvers re-fit on the noisy input (that is their inference). The ground truth stays the clean phantom, and the FBP baseline is recomputed from the noisy sinogram, so hr = max(0, 1 − rmse / rmse_noisy-FBP) measures improvement over the noisy FBP. This is exactly the realistic M_phys = M + ε extension the Sidky challenge flags but does not study.

Leaderboard — every solver (rendered from the registry)

Rendered live from docs/runs/index/leaderboard.json (build_registry.py), same run-ids as Breast-CT but scored on the noised inputs. hr is fair to rank within this board (all solvers share the same noisy FBP baseline); the absolute SSIM / PSNR columns (vs the clean truth) distinguish genuine robustness from baseline-gaming.

The finding — ideal-data performance does not predict noise robustness

The ranking nearly inverts relative to the noiseless board:

• learned-primal-dual becomes champion (hr 0.72 → 0.93, SSIM 0.9905) — its primal-dual data-consistency + learned regularization is genuinely noise-robust.
• Bilateral / self-supervised / classical methods rise: manduca-bilateral (21 params) 0.28 → 0.84, dual-domain-bilateral-n2i → 0.79, manhart-PWLS-TV and tv-iterative 0.36 → 0.51 — explicit or implicit smoothing suppresses noise.
• The noiseless champion collapses to last. dual-domain-supervised, a pure supervised streak-remover, drops 0.89 → 0.00 (SSIM 0.9992 → 0.35): trained only on noiseless FBP, it is brittle to a distribution it never saw. The fastdiff constrained variants collapse likewise.
• DC-unrolled networks (ItNet family) degrade moderately (0.89 → 0.55–0.70): the data-consistency step re-injects some measurement noise, but the learned prior partly holds. param-efficient (195 params) holds relatively well (0.62 → 0.52) — its output bilateral filter helps.

Take-away for the paper: a leaderboard won under ideal conditions says little about deployment robustness. The methods that top the noiseless board (supervised denoisers optimizing exact inversion) are the least robust; the physically regularized methods (primal-dual, bilateral, TV) that sit mid-pack on ideal data are the most robust. Report both boards.

DNF. naf and r2gaussian (per-scene neural-implicit / Gaussian-splatting) are included but do not complete — per-scene optimization scores ~20/200 cases even with a 5-hour wall and hr ≈ 0 on those, on noisy as on noiseless data.

Methodology

Noise staging: data/stage_breast_noise.py (deterministic, seed 20260709). No-retrain eval: scripts/score_breast_noise.py → score_breast_testset.py --noise-i0. See paper plan §5.6.7.