Breast-CT leaderboard

128-view 2-D fan-beam sparse-view synthetic phantoms (Sidky-group breast model, real μ range up to ~0.5). All metrics through ddssl_ldct.metrics.evaluate_calibrated: two-point linear intensity calibration on the foreground inside an inscribed-circle FoV mask, then PSNR/SSIM/RMSE on the calibrated prediction.

Baseline FBP: SSIM = 0.957, PSNR = 39.74 dB, hr = 0. hr = max(0, 1 − rmse / baseline_rmse).

Leaderboard — every solver (rendered from the registry)

The table below is rendered live from the registry (docs/runs/index/leaderboard.json, built by scripts/build_registry.py from each iter’s immutable observation.json). It lists every solver exercised on breast-CT, ranked by val headroom (SSIM tiebreak). Breast-CT has no held-out test set, so every column is on the val set — SSIM / PSNR / RMSE are shown as mean ± std over the scored slices. The below-baseline / discarded solvers (hr = 0, structurally bounded) stay on the board dimmed and unranked so the full inventory always shows — never a top-N. No number on this page is typed by hand. TPE runs that predate the PSNR/RMSE/time/params logging show for those columns (only SSIM + headroom were recorded then, and the raw recon is not retained). params (M) is trainable parameters in millions (a raw integer for the handful-of-params solvers); 0.000/small values are exact.

All metrics use ddssl_ldct.metrics.evaluate_calibrated: two-point linear intensity calibration on the foreground inside an inscribed-circle FoV mask, then PSNR/SSIM/RMSE on the calibrated prediction. Baseline FBP: SSIM = 0.957, PSNR = 39.74 dB, hr = 0.

loading leaderboard…

The dimmed (hr = 0) rows are not “just under the threshold” — they are structurally bounded by FBP under the calibrated metric, as detailed below.

Why these fail structurally

Inventory-gap + Wu-TPE closure log (2026-06-09)

ItNet v1 was dispatched + Wu 2015 trainable got a full TPE pass on breast-CT on 2026-06-09:

Solver TPE job Result Status
ItNet v1 (solver_itnet.py) 762956 hr=0.1703 at iter-20 winner (k=8, c=8, pretrain_ep=5, lr=1e-3, α=0.015) ABOVE BASELINE — slots in at rank 15. Surprises the Mayo verdict (hr=0): on breast-CT’s broader synthetic-anatomy substrate, v1’s deeper finetune-pass lets it clear baseline. Confirms ItNet family transfer pattern: v1 < v2 < v3 on breast-CT (0.1703 → 0.5386 → 0.7342).
Wu 2015 trainable (solver_wu_2015_trainable.py) 762955 COMPLETE 20/20: final hr=0.3170 (winner cluster across iter-8/12/13/15/16 at hr=0.31-0.32, all with n_bands=6, n_outer=2, range=8, window=2, ep~12-13, lr~1e-4, λ_neg~0.7) 🎯 +45% over agentic 0.2189! TPE found a higher-n_bands + lower-lr corner the agentic search missed. FINAL at rank 10. Agentic-seed row stays at rank 14.

Cross-dataset coverage compare (after 2026-06-09 gap-closure dispatches):

Methodology

See solver_plan.md. One row per solver family; “Variant” picks the best config across all autoresearch + TPE iterations.