| """Templated per-tile text for ORION (no real text is paired with ORION tiles). |
| |
| Each tile gets 4 paraphrases describing cellularity and cell-type composition derived from the |
| marker-positive nuclei counts (fraction of all nuclei) — i.e. the text carries a coarse, discretised |
| version of the marker information. Kept short (< 77 CLIP tokens). |
| Output: cache/text.json {tile_id: [4 strings]} |
| """ |
| import json |
| import os |
|
|
| import numpy as np |
| import pandas as pd |
|
|
| from common import CACHE, ROOT, load_all_tiles |
|
|
| df = load_all_tiles() |
| stats = pd.read_csv(os.path.join(ROOT, "results", "tile_stats.csv")).set_index("tile_id") |
| df["n_nuclei"] = stats.loc[df.tile_id, "n_nuclei"].values |
|
|
| CELLS = [ |
| ("Pan-CK_count", "Pan-CK+ epithelial cells"), ("SMA_count", "SMA+ stromal cells"), |
| ("CD45_count", "CD45+ immune cells"), ("CD3e_count", "CD3e+ T cells"), ("CD8a_count", "CD8a+ cytotoxic T cells"), |
| ("CD4_count", "CD4+ helper T cells"), ("FOXP3_count", "FOXP3+ regulatory T cells"), ("CD20_count", "CD20+ B cells"), |
| ("CD68_count", "CD68+ macrophages"), ("CD163_count", "CD163+ macrophages"), ("CD31_count", "CD31+ endothelial cells"), |
| ("Ki67_count", "Ki67+ proliferating cells"), ("PD-L1_count", "PD-L1+ cells"), ("PD-1_count", "PD-1+ cells"), |
| ] |
| LEVEL_WORDS = [ |
| ("few", "moderate", "abundant"), ("scattered", "some", "many"), |
| ("rare", "a moderate number of", "numerous"), ("sparse", "intermediate", "dense"), |
| ] |
| DENSITY = [("acellular", "sparsely cellular", "moderately cellular", "highly cellular"), |
| ("background", "low-density", "medium-density", "dense"), |
| ("nearly empty", "sparse", "moderately dense", "crowded"), |
| ("mostly non-tissue", "hypocellular", "cellular", "hypercellular")] |
| OPEN = ["H&E tile, {d} region with {n} nuclei.", "A {d} tissue patch ({n} nuclei).", |
| "Colorectal cancer tissue, {d}, about {n} nuclei.", "{D} area containing {n} nuclei."] |
|
|
|
|
| def level(frac): |
| return None if frac < 0.02 else 0 if frac < 0.08 else 1 if frac < 0.25 else 2 |
|
|
|
|
| def describe(r, v, rng): |
| n = int(r.n_nuclei) |
| di = 0 if n < 20 else 1 if n < 100 else 2 if n < 250 else 3 |
| d = DENSITY[v][di] |
| n_round = int(round(n, -1)) |
| parts = [OPEN[v].format(d=d, D=d.capitalize(), n=n_round)] |
| if n >= 20: |
| items = [] |
| for col, name in CELLS: |
| lv = level(r[col] / max(n, 1)) |
| if lv is not None: |
| items.append((lv, f"{LEVEL_WORDS[v][lv]} {name}")) |
| items.sort(key=lambda x: -x[0]) |
| items = [s for _, s in items[:6]] |
| if v % 2 == 1: |
| rng.shuffle(items) |
| parts.append(("Shows " + ", ".join(items) + ".") if items else "No marker-positive population stands out.") |
| return " ".join(parts) |
|
|
|
|
| rng = np.random.default_rng(0) |
| texts = {int(r.tile_id): [describe(r, v, rng) for v in range(4)] for _, r in df.iterrows()} |
| json.dump(texts, open(os.path.join(CACHE, "text.json"), "w"), indent=0) |
| u = len(set(t[0] for t in texts.values())) |
| print(f"{len(texts)} tiles, {u} unique variant-0 texts") |
| for k in [0, 500, 1500]: |
| print(texts[k]) |
|
|
