BreastCT-Noise-Retrained leaderboard — matched-noise retraining

The BreastCT-Noise board is a no-retrain robustness test: clean-optimal models are re-evaluated on noisy inputs they never trained on, and the ideal-data ranking nearly inverts. This third board asks the converse question: if the noise is known at training time, does the clean ranking return?

Setup (matched-noise retraining). For each trainable solver we take its noiseless best-iteration configuration — unchanged architecture and hyper-parameters — and retrain it from scratch on the same Breast-CT training set with matched Poisson noise added (I0 = 100 000 photons/pixel, deterministic seed), select the checkpoint on the noisy validation split, and score it on the same 200 noisy held-out test cases as the no-retrain board. The ground truth stays clean and the FBP baseline is recomputed from the noisy sinogram, so the retrained and no-retrain boards share the same hr scale and are directly comparable. Classical / per-scene / zero-shot solvers carry no trainable weights and are carried over unchanged from the no-retrain board. Only the frozen clean-best config was retrained (no hyper-parameter re-search), so these scores are a conservative lower bound.

Leaderboard — every solver (rendered from the registry)

Rendered live from docs/runs/index/leaderboard.json (build_registry.py / build_breast_retrain_board.py), same run-ids as Breast-CT but scored after matched-noise retraining.

The finding — the collapse is a distribution mismatch, not a deficit

Retraining on matched noise largely restores the clean ranking. Its Spearman correlation with the noiseless board rises from ρ ≈ 0.16 (no retrain, ~3 % of the variance) to ρ ≈ 0.65 (retrained, ~42 %).

• The collapsed noiseless champion recovers to #1. dual-domain-supervised goes hr 0.00 → 0.958 (SSIM 0.35 → 0.996) and reclaims rank 1 — above its own noiseless score, because hr is measured against the worse noisy FBP baseline. Its collapse was a train/test distribution mismatch, not an architecture limit.
• The supervised unrolled family returns to the top tier (ITNet v1/v2/v3, U-Swin, hr ≈ 0.94).
• learned-primal-dual barely moves (0.935 → 0.936): already robust because it keeps the forward operator in the loop, it had nothing to regain.
• Two groups do not recover: the constrained fast-diffusion variants stay at hr 0 (their hard fit to the clean forward model does not survive noisy training), and a few solvers dip slightly (e.g. Hammernik-2017, 0.70 → 0.61) because only the fixed clean-best config was retrained, not re-searched.

Take-away for the paper: a single-distribution leaderboard does not predict robustness to unseen noise — but that brittleness is largely repairable once the noise is known and trained for. A complete benchmark should report both conditions: transfer to an unseen perturbation, and performance after matched retraining.

Methodology

Noise staging (train + val + test): data/stage_breast_noise.py (deterministic, seed 20260709). Retraining + scoring: scripts/score_breast_noise_retrain.py → score_breast_testset.py --worker --noise-i0 100000 --retrain. Board assembly + rank correlation: scripts/build_breast_retrain_board.py and scripts/breast_board_rank_corr.py.