"""Why GigaTIME virtual mIF fails at instance-level pairing (ORION, 10 shared channels, same tiles as the real mIF). (a) pooled vs WITHIN-slide Spearman of tile means: pairing in a 100-tile gallery needs tile-to-tile differences inside a slide, while pooled correlation is dominated by slide/tissue-level differences (b) direct real-mean <-> virtual-mean retrieval (no learned head): rank-gaussianised channel means, cosine, g100 (c) in-domain reference: ridge UNI2 H&E -> real means, trained on ORION patient_5fold train, scored on test (what H&E can predict about these 10 channels when the model has seen ORION) (d) input distribution: real mIF intensities vs GigaTIME sigmoid probabilities (what the VirTues head receives) Output results/gigatime_diag.csv + printed tables. """ import json import os import numpy as np import pandas as pd from scipy.special import ndtri from scipy.stats import rankdata, spearmanr from common import CACHE, ROOT, SPLITS, load_all_tiles OUT = os.path.join(ROOT, "results") G = os.path.join(CACHE, "gigatime") ORION_CH = ["Hoechst", "CD31", "CD45", "CD68", "CD4", "FOXP3", "CD8a", "CD45RO", "CD20", "PD-L1", "CD3e", "CD163", "E-cadherin", "PD-1", "Ki67", "Pan-CK", "SMA"] PANEL = [("Hoechst", ["DAPI"]), ("CD68", ["CD68"]), ("CD4", ["CD4"]), ("CD8a", ["CD8"]), ("CD20", ["CD20"]), ("PD-L1", ["PD-L1"]), ("CD3e", ["CD3"]), ("Ki67", ["Ki67"]), ("Pan-CK", ["CK"]), ("SMA", ["Actin-D", "Transgelin"])] gch = json.load(open(os.path.join(G, "channels.json")))["channels"] df = load_all_tiles() slide = df.in_slide_name.values real = np.load(os.path.join(G, "orion_real_if_means.npy"))[:, [ORION_CH.index(o) for o, _ in PANEL]] vm = np.load(os.path.join(G, "orion_means.npy")) virt = np.stack([vm[:, [gch.index(g) for g in gs]].mean(1) for _, gs in PANEL], 1) he = np.load(os.path.join(CACHE, "he_uni2.npy"))[:, 0].astype(np.float64) names = [o for o, _ in PANEL] rows = [] def within(a, b): """mean within-slide Spearman (slides with >= 20 tiles), tiles-weighted""" r, w = [], [] for s in np.unique(slide): m = slide == s if m.sum() >= 20 and a[m].std() > 0 and b[m].std() > 0: r.append(spearmanr(a[m], b[m])[0]); w.append(m.sum()) return float(np.average(r, weights=w)) rg = lambda x: ndtri((np.apply_along_axis(rankdata, 0, x) - 0.5) / len(x)) def g100(a, b, draws=50, seed=0, groups=None): a = a / np.linalg.norm(a, axis=1, keepdims=True); b = b / np.linalg.norm(b, axis=1, keepdims=True) rng = np.random.default_rng(seed); r = [] for _ in range(draws): if groups is None: s = rng.choice(len(a), min(100, len(a)), replace=False) else: # gallery drawn from ONE slide: only within-slide differences can help g = rng.choice(groups); s = np.flatnonzero(slide == g); s = rng.choice(s, min(100, len(s)), replace=False) sim = a[s] @ b[s].T; d = np.diag(sim)[:, None] r.append(((sim > d).sum(1) == 0).mean() / 2 + ((sim.T > d).sum(1) == 0).mean() / 2) return float(np.mean(r)) # ---- (c) ridge UNI2 -> real means (in-domain reference), patient 5-fold test predictions pred = np.zeros_like(real, dtype=np.float64) R_rg = rg(real) for k in range(5): s = pd.read_csv(os.path.join(SPLITS, "patient_5fold", f"fold{k}.csv")) tr, va, te = (s.tile_id[s.split == x].values for x in ["train", "val", "test"]) mu, sd = he[tr].mean(0), he[tr].std(0) + 1e-6 X = lambda i: (he[i] - mu) / sd best = None for lam in [1e1, 1e2, 1e3, 1e4]: # lambda chosen on val W = np.linalg.solve(X(tr).T @ X(tr) + lam * np.eye(he.shape[1]), X(tr).T @ R_rg[tr]) v = np.mean([spearmanr(X(va) @ W[:, c], R_rg[va, c])[0] for c in range(len(names))]) if best is None or v > best[0]: best = (v, W) pred[te] = X(te) @ best[1] # ---- (a) correlations for c, n in enumerate(names): rows.append({"channel": n, "GigaTIME pooled": spearmanr(virt[:, c], real[:, c])[0], "GigaTIME within-slide": within(virt[:, c], real[:, c]), "UNI2-ridge pooled (test)": spearmanr(pred[:, c], real[:, c])[0], "UNI2-ridge within-slide (test)": within(pred[:, c], real[:, c])}) C = pd.DataFrame(rows).set_index("channel") C.loc["mean"] = C.mean() # ---- (b) direct retrieval, all tiles, galleries mixed across slides vs within one slide big = [s for s in np.unique(slide) if (slide == s).sum() >= 100] V_rg, P_rg = rg(virt), rg(pred) Rt = pd.DataFrame([ {"pair": "real means <-> GigaTIME means", "mixed-slide gallery": g100(R_rg, V_rg), "single-slide gallery": g100(R_rg, V_rg, groups=big)}, {"pair": "real means <-> UNI2-ridge prediction", "mixed-slide gallery": g100(R_rg, P_rg), "single-slide gallery": g100(R_rg, P_rg, groups=big)}, {"pair": "real means <-> shuffled real (chance)", "mixed-slide gallery": g100(R_rg, R_rg[np.random.default_rng(0).permutation(len(R_rg))]), "single-slide gallery": np.nan}, ]).set_index("pair") # ---- (d) input distribution of the pixel maps (200 tiles) maps = np.load(os.path.join(G, "orion_maps.npy"), mmap_mode="r") idx = np.random.default_rng(0).choice(len(maps), 200, replace=False) mv = np.asarray(maps[np.sort(idx)])[:, [gch.index(gs[0]) for _, gs in PANEL]].astype(np.float32) / 255 D = pd.DataFrame({"channel": names, "virtual: frac px p<0.05": (mv < 0.05).mean((0, 2, 3)), "virtual: frac px p>0.95": (mv > 0.95).mean((0, 2, 3)), "virtual: frac px in (0.05,0.95)": ((mv >= 0.05) & (mv <= 0.95)).mean((0, 2, 3))}).set_index("channel") pd.set_option("display.width", 200) print("(a) Spearman of tile means vs real ORION mIF\n", C.round(3).to_string()) print("\n(b) R@1 in 100-tile galleries (chance 0.01), no learned head\n", Rt.round(3).to_string()) print("\n(d) GigaTIME pixel-value distribution (sigmoid probabilities)\n", D.round(3).to_string()) C.round(4).to_csv(os.path.join(OUT, "gigatime_diag.csv")) Rt.round(4).to_csv(os.path.join(OUT, "gigatime_diag_retrieval.csv")) D.round(4).to_csv(os.path.join(OUT, "gigatime_diag_pixels.csv"))