Data & reproducibility

# §5.1 — no significant DistMap Dice gain at convergence
import json, numpy as np
from scipy.stats import wilcoxon
d = json.load(open('rankings.json'))
for reg in ('WT', 'TC', 'ET'):
    B = np.array([r['baseline'][f'dice_{reg}'] for r in d['rows']])
    D = np.array([r['distmap'][f'dice_{reg}'] for r in d['rows']])
    _, p = wilcoxon(D, B, alternative='greater')
    print(f"{reg}: Δ={100*(D-B).mean():+.3f} pp, p={p:.2e}")

# §5.2 — topological fragment reduction
import csv
from scipy.stats import wilcoxon
rows = list(csv.DictReader(open('fragments_topological_cv.csv')))
for cls in ('NCR', 'ED', 'ET'):
    D = np.array([int(r[f'frags_D_{cls}']) for r in rows])
    F = np.array([int(r[f'frags_F_{cls}']) for r in rows])
    _, p = wilcoxon(F, D, alternative='less')
    print(f"{cls}: Δ={100*(F.mean()-D.mean())/D.mean():+.1f}% vs DistMap, p={p:.2e}")

# §5.3 — per-class HD95 NCR
rows = list(csv.DictReader(open('hd95_per_class_cv.csv')))
F = np.array([float(r['hd95_NCR_F']) for r in rows if r['hd95_NCR_F']])
D = np.array([float(r['hd95_NCR_D']) for r in rows if r['hd95_NCR_D']])
_, p = wilcoxon(F, D, alternative='less')
print(f"NCR HD95: F={F.mean():.3f} vs D={D.mean():.3f}, p={p:.2e}")