| """PathGen-1.6M H&E-caption pairs (real text) -> cache/pathgen_meta.csv with a slide-level 80/10/10 split. |
| Used to bind the text modality to the H&E anchor (ImageBind: text only needs to be paired with images).""" |
| import json |
| import os |
|
|
| import numpy as np |
| import pandas as pd |
|
|
| from common import CACHE, ROOT |
|
|
| PG = os.path.join(ROOT, "..", "PathGen-1.6M") |
| d = json.load(open(os.path.join(PG, "Data", "PathGen-1.6M.json"))) |
| m = pd.DataFrame([{"wsi_id": x["wsi_id"], "caption": x["caption"], |
| "image_path": os.path.join(PG, "Data", "images", f"{x['wsi_id']}_{x['position'][0]}_{x['position'][1]}.png")} |
| for x in d]) |
| assert m.image_path.map(os.path.exists).all() |
| m["patient"] = m.wsi_id.str[:12] |
| pats = np.random.default_rng(0).permutation(sorted(m.patient.unique())) |
| n = len(pats) |
| role = {p: "train" if i < 0.8 * n else "val" if i < 0.9 * n else "test" for i, p in enumerate(pats)} |
| m["split"] = m.patient.map(role) |
| m["project_code"] = m.wsi_id.str[5:7] |
| m.to_csv(os.path.join(CACHE, "pathgen_meta.csv"), index_label="idx") |
| print(m.split.value_counts().to_dict(), "patients:", n) |
|
|
