| """Step 0: convert each open source's native annotation files into <raw>/<source>/labels.csv |
| |
| Columns: image (path relative to the source dir), label ('|'-separated native tokens), patient, eye, grade (ICDR 0-4) |
| Native tokens are written verbatim (prefixed with the column / scale they come from) so nothing is lost. |
| python -m fsbench.prepare_labels --raw data/raw [--only hygd grape ...] |
| """ |
| import argparse |
| import re |
| from pathlib import Path |
|
|
| import pandas as pd |
|
|
| IMG = {".jpg", ".jpeg", ".png", ".tif", ".tiff", ".bmp"} |
|
|
|
|
| def _imgs(d): |
| return sorted(p for p in Path(d).rglob("*") if p.suffix.lower() in IMG and not p.name.startswith("._")) |
|
|
|
|
| def _rows_to_df(rows): |
| return pd.DataFrame(rows, columns=["image", "label", "patient", "eye", "grade"]) |
|
|
|
|
| def hrf(src): |
| tag = {"h": "healthy", "dr": "diabetic_retinopathy", "g": "glaucoma"} |
| rows = [] |
| for p in _imgs(src / "images"): |
| m = re.match(r"(\d+)_(h|dr|g)\.", p.name, re.I) |
| rows.append([str(p.relative_to(src)), tag[m.group(2).lower()], "", "", ""]) |
| return _rows_to_df(rows) |
|
|
|
|
| def hygd(src): |
| d = pd.read_csv(src / "Labels.csv").dropna(axis=1, how="all") |
| return _rows_to_df([[f"Images/{r['Image Name']}", r["Label"], str(r["Patient"]), "", ""] for _, r in d.iterrows()]) |
|
|
|
|
| def grape(src): |
| rows = [] |
| for p in _imgs(src / "CFPs"): |
| pid, eye, visit = p.stem.split("_")[:3] |
| rows.append([str(p.relative_to(src)), "glaucoma_follow_up", pid, {"OD": "right", "OS": "left"}.get(eye, eye), ""]) |
| return _rows_to_df(rows) |
|
|
|
|
| def sustech_sysu(src): |
| d = pd.read_csv(src / "originalImages/drLabels.csv") |
| c5 = pd.read_csv(src / "originalImages/c5_DR_reclassified.csv").set_index("Fundus_images").iloc[:, 0].to_dict() |
| rows = [] |
| for _, r in d.iterrows(): |
| g = int(r.iloc[2]) |
| icdr = c5.get(r.iloc[0], "") if g == 5 else g |
| tok = f"icdr_{g}" if g < 5 else f"laser_scars(reclassified_icdr_{icdr})" |
| rows.append([f"originalImages/{r.iloc[0]}", tok, "", ["left", "right"][int(r.iloc[1])], str(icdr)]) |
| return _rows_to_df(rows) |
|
|
|
|
| def deepdrid(src): |
| base = src / "DeepDRiD-master/regular_fundus_images" |
| rows = [] |
| for part in ("training", "validation"): |
| d = pd.read_csv(base / f"regular-fundus-{part}/regular-fundus-{part}.csv") |
| for _, r in d.iterrows(): |
| g = r["left_eye_DR_Level"] if pd.notna(r["left_eye_DR_Level"]) else r["right_eye_DR_Level"] |
| rel = f"DeepDRiD-master/regular_fundus_images/regular-fundus-{part}/Images/{r['patient_id']}/{r['image_id']}.jpg" |
| q = "good_quality" if int(r["Overall quality"]) == 1 else "poor_quality" |
| rows.append([rel, f"icdr_{int(g)}|{q}", str(r["patient_id"]), "left" if "_l" in r["image_id"] else "right", str(int(g))]) |
| t = pd.read_excel(base / "Online-Challenge1&2-Evaluation/Challenge1_labels.xlsx") |
| for _, r in t.iterrows(): |
| pid = r["image_id"].split("_")[0] |
| rel = f"DeepDRiD-master/regular_fundus_images/Online-Challenge1&2-Evaluation/Images/{pid}/{r['image_id']}.jpg" |
| rows.append([rel, f"icdr_{int(r['DR_Levels'])}", pid, "left" if "_l" in r["image_id"] else "right", str(int(r["DR_Levels"]))]) |
| return _rows_to_df(rows) |
|
|
|
|
| def rfmid2(src): |
| rows = [] |
| for part, f in (("Training_set", "RFMiD_2_Training_labels.csv"), ("Validation_set", "RFMiD_2_Validation_labels.csv"), |
| ("Test_set", "RFMiD_2_Testing_labels.csv")): |
| d = pd.read_csv(src / part / f, encoding="latin-1") |
| d = d.loc[:, ~d.columns.str.startswith("Unnamed")] |
| cols = [c for c in d.columns if c != "ID"] |
| for _, r in d.iterrows(): |
| toks = [c.strip() for c in cols if r[c] == 1] |
| img = next((p for p in (src / part).glob(f"{r['ID']}.*") if p.suffix.lower() in IMG), None) |
| rows.append([str(img.relative_to(src)) if img else f"{part}/{r['ID']}.jpg", "|".join(toks), "", "", ""]) |
| return _rows_to_df(rows) |
|
|
|
|
| def idrid(src): |
| base = "B. Disease Grading" |
| rows = [] |
| for gt, imgdir in (("a. IDRiD_Disease Grading_Training Labels.csv", "1. Original Images/a. Training Set"), |
| ("b. IDRiD_Disease Grading_Testing Labels.csv", "1. Original Images/b. Testing Set")): |
| d = pd.read_csv(src / base / "2. Groundtruths" / gt).dropna(axis=1, how="all") |
| d.columns = [c.strip() for c in d.columns] |
| for _, r in d.iterrows(): |
| g = int(r["Retinopathy grade"]) |
| rows.append([f"{base}/{imgdir}/{r['Image name']}.jpg", f"icdr_{g}|dme_risk_{int(r['Risk of macular edema'])}", |
| "", "", str(g)]) |
| return _rows_to_df(rows) |
|
|
|
|
| def papila(src): |
| base = next(src.glob("PapilaDB*")) |
| diag = {0: "healthy", 1: "glaucoma", 2: "suspect"} |
| rows = [] |
| for eye, f in (("right", "patient_data_od.xlsx"), ("left", "patient_data_os.xlsx")): |
| d = pd.read_excel(base / "ClinicalData" / f, header=None).iloc[3:] |
| for _, r in d.iterrows(): |
| pid = str(r[0]).lstrip("#") |
| img = base / "FundusImages" / f"RET{pid}{'OD' if eye == 'right' else 'OS'}.jpg" |
| rows.append([str(img.relative_to(src)), diag[int(r[3])], pid, eye, ""]) |
| return _rows_to_df(rows) |
|
|
|
|
| def jichi(src): |
| d = pd.read_csv(src / "dmr/list.csv") |
| rows = [] |
| for _, r in d.iterrows(): |
| pid, _, eye = Path(r["Image"]).stem.split("_") |
| rows.append([f"dmr/{r['Image']}", f"davis_{r['Davis_grading_of_one_figure'].strip().lower()}", pid, |
| {"R": "right", "L": "left"}[eye], ""]) |
| return _rows_to_df(rows) |
|
|
|
|
| def fives(src): |
| tag = {"A": "amd", "D": "diabetic_retinopathy", "G": "glaucoma", "N": "normal"} |
| rows = [] |
| for p in _imgs(src): |
| m = re.fullmatch(r"\d+_([ADGN])", p.stem) |
| if "Original" in p.parts and m: |
| rows.append([str(p.relative_to(src)), tag[m.group(1)], "", "", ""]) |
| return _rows_to_df(rows) |
|
|
|
|
| def _class_folders(src, base): |
| """One class per sub-folder; token = folder name with spaces -> '_' (no patient information released).""" |
| return _rows_to_df([[str(p.relative_to(src)), p.parent.name.replace(" ", "_"), "", "", ""] |
| for p in _imgs(src / base)]) |
|
|
|
|
| def bangladesh_eye(src): |
| return _class_folders(src, "Original Dataset") |
|
|
|
|
| def paraguay_dr(src): |
| """7-level Paraguay scale -> ICDR: 1 none, 2 mild, 3 moderate, 4/5 severe & very severe NPDR, 6/7 PDR.""" |
| d = _class_folders(src, "") |
| icdr = {"1": "0", "2": "1", "3": "2", "4": "3", "5": "3", "6": "4", "7": "4"} |
| d["grade"] = [icdr.get(t.split(".")[0], "") for t in d.label] |
| return d |
|
|
|
|
| def chaksu(src): |
| """Majority decision of 5 experts per device; native token keeps the vote, e.g. glaucoma_suspect|votes_3of5.""" |
| rows = [] |
| for part in ("Train", "Test"): |
| dec = src / part / "6.0_Glaucoma_Decision" |
| for dev in ("Bosch", "Forus", "Remidio"): |
| f = dec / f"Glaucoma_Decision_Comparison_{dev}_majority.csv" |
| if not f.exists(): |
| continue |
| d = pd.read_csv(f, encoding="utf-8-sig") |
| imgs = {p.stem: p for p in _imgs(src / part / "1.0_Original_Fundus_Images" / dev)} |
| experts = [c for c in d.columns if c.lower().startswith("expert")] |
| for _, r in d.iterrows(): |
| stem = r["Images"].split(".")[0] |
| if stem not in imgs: |
| continue |
| norm = lambda v: re.sub(r"\s+", "_", str(v).strip().lower()).replace("susupect", "suspect") |
| maj = norm(r["Glaucoma Decision"] if "Glaucoma Decision" in d.columns else r["Majority Decision"]) |
| votes = sum(norm(r[c]) == maj for c in experts) |
| m = re.match(r"(P\d+)_", stem) |
| rows.append([str(imgs[stem].relative_to(src)), f"{maj}|votes_{votes}of{len(experts)}|device_{dev}", |
| f"{dev}:{m.group(1)}" if m else "", "", ""]) |
| return _rows_to_df(rows) |
|
|
|
|
| def palm(src): |
| """PALM: Label 1 = pathologic myopia. Only the training part names non-PM images as N (normal) or H (high myopia); |
| validation/testing non-PM images may be either, so they are kept as 'non_pm' (excluded from N).""" |
| rows = [] |
| for part in ("Training", "Validation", "Testing"): |
| d = pd.read_excel(src / "PALM" / part / "Classification Labels.xlsx") |
| sup = pd.read_excel(src / "PALM" / part / "Supplementary Information.xlsx").set_index("imgName").iloc[:, 0] |
| for _, r in d.iterrows(): |
| name = r["imgName"] |
| if int(r["Label"]) == 1: |
| tok = "pathologic_myopia" |
| else: |
| tok = {"N": "normal", "H": "high_myopia_no_pm"}.get(name[0], "non_pm") |
| dev = {1: "Zeiss", 2: "Canon"}.get(int(sup.get(name, 0)), "") |
| rows.append([f"PALM/{part}/Images/{name}", f"{tok}|device_{dev}", "", "", ""]) |
| return _rows_to_df(rows) |
|
|
|
|
| def afio_composite(src): |
| """AFIO (Rawalpindi) composite fundus+OCT set: colour fundus photographs only; class = folder, patient = MRN prefix.""" |
| base = src / "Dataset/Dataset" |
| rows = [] |
| for p in _imgs(base): |
| if "_Color_" not in p.name: |
| continue |
| group, cls = p.relative_to(base).parts[:2] |
| eye = "right" if "_R_" in p.name else "left" |
| rows.append([str(p.relative_to(src)), f"{group.lower()}_{cls.replace(' ', '_').lower()}", p.name.split("_")[0], eye, ""]) |
| return _rows_to_df(rows) |
|
|
|
|
| def ddr(src): |
| """OIA-DDR DR grading split files: '<image> <grade>' (0-4 ICDR, 5 = ungradable).""" |
| base = next(src.rglob("DR_grading")) |
| rows = [] |
| for part in ("train", "valid", "test"): |
| d = pd.read_csv(base / f"{part}.txt", sep=" ", header=None, names=["image", "grade"], dtype=str) |
| for _, r in d.iterrows(): |
| g = r["grade"].strip() |
| rows.append([str((base / part / r["image"]).relative_to(src)), f"ddr_{g}" if g != "5" else "ungradable", |
| "", "", g if g != "5" else ""]) |
| return _rows_to_df(rows) |
|
|
|
|
| PREPARERS = {"ddr": ddr, "palm": palm, "afio_composite": afio_composite, "chaksu": chaksu, "idrid": idrid, "papila": papila, "jichi": jichi, "fives": fives, "bangladesh_eye": bangladesh_eye, |
| "paraguay_dr": paraguay_dr, "hrf": hrf, "hygd": hygd, "grape": grape, "sustech_sysu": sustech_sysu, "deepdrid": deepdrid, |
| "rfmid2": rfmid2} |
|
|
|
|
| def main(argv=None): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--raw", required=True) |
| ap.add_argument("--only", nargs="*") |
| a = ap.parse_args(argv) |
| for sid, fn in PREPARERS.items(): |
| src = Path(a.raw) / sid |
| if (a.only and sid not in a.only) or not src.is_dir(): |
| continue |
| df = fn(src) |
| df.to_csv(src / "labels.csv", index=False) |
| miss = sum(not (src / p).is_file() for p in df.image) |
| print(f"{sid:<14} rows={len(df):>6} missing={miss:>4} top labels={df.label.value_counts().head(6).to_dict()}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
