| import os |
| import io |
| from pathlib import Path |
|
|
| import pandas as pd |
| import numpy as np |
| from PIL import Image |
|
|
|
|
| INPUT_PARQUET_PATHS = [ |
| r"./data/fold1-00000-of-00001.parquet", |
| r"./data/fold2-00000-of-00001.parquet", |
| r"./data/fold3-00000-of-00001.parquet", |
| ] |
|
|
| SPLIT_CSV_PATH = r"./PanNuke_split.csv" |
|
|
| OUTPUT_TRAIN_ROOT = Path(r"./train_images") |
| OUTPUT_TEST_ROOT = Path(r"./test_images") |
|
|
| IMAGE_COLUMN = "image" |
| CLASS_LABEL_COLUMN = "tissue" |
| IMAGE_EXT = ".png" |
|
|
| TISSUE_ID_TO_NAME: dict[int, str] = { |
| 0: "Adrenal Gland", |
| 1: "Bile Duct", |
| 2: "Bladder", |
| 3: "Breast", |
| 4: "Cervix", |
| 5: "Colon", |
| 6: "Esophagus", |
| 7: "Head and Neck", |
| 8: "Kidney", |
| 9: "Liver", |
| 10: "Lung", |
| 11: "Ovarian", |
| 12: "Pancreatic", |
| 13: "Prostate", |
| 14: "Skin", |
| 15: "Stomach", |
| 16: "Testis", |
| 17: "Thyroid", |
| 18: "Uterus", |
| } |
|
|
| CLASS_ORDER = list(TISSUE_ID_TO_NAME.values()) |
|
|
|
|
| def load_split_mapping(csv_path: str) -> dict[str, str]: |
| df = pd.read_csv(csv_path) |
| df.columns = df.columns.str.strip() |
| mapping = {} |
| for _, row in df.iterrows(): |
| img_path = str(row["image_path"]).strip() |
| split = str(row["split"]).strip().lower() |
| mapping[img_path] = split |
| print(f"[INFO] Loaded split CSV: {len(mapping)} entries " |
| f"(train: {sum(1 for v in mapping.values() if v == 'train')}, " |
| f"test: {sum(1 for v in mapping.values() if v == 'test')})") |
| return mapping |
|
|
|
|
| def is_null_like(x): |
| if x is None: |
| return True |
| if isinstance(x, (bytes, bytearray, dict, list, tuple, np.ndarray, Image.Image)): |
| return False |
| try: |
| result = pd.isna(x) |
| if isinstance(result, (bool, np.bool_)): |
| return bool(result) |
| return False |
| except Exception: |
| return False |
|
|
|
|
| def try_decode_pil_from_bytes(data): |
| try: |
| img = Image.open(io.BytesIO(data)) |
| img.load() |
| return img |
| except Exception: |
| return None |
|
|
|
|
| def try_decode_image(obj): |
| if is_null_like(obj): |
| return None |
| if isinstance(obj, Image.Image): |
| return obj |
| if isinstance(obj, (bytes, bytearray)): |
| return try_decode_pil_from_bytes(obj) |
| if isinstance(obj, np.ndarray): |
| try: |
| arr = obj.astype(np.uint8) if obj.dtype != np.uint8 else obj |
| if arr.ndim in (2, 3): |
| return Image.fromarray(arr) |
| except Exception: |
| return None |
| if isinstance(obj, list): |
| try: |
| arr = np.array(obj) |
| arr = arr.astype(np.uint8) if arr.dtype != np.uint8 else arr |
| if arr.ndim in (2, 3): |
| return Image.fromarray(arr) |
| except Exception: |
| return None |
| if isinstance(obj, dict): |
| if "bytes" in obj and obj["bytes"] is not None: |
| return try_decode_pil_from_bytes(obj["bytes"]) |
| if "array" in obj and obj["array"] is not None: |
| try: |
| arr = np.array(obj["array"]) |
| arr = arr.astype(np.uint8) if arr.dtype != np.uint8 else arr |
| if arr.ndim in (2, 3): |
| return Image.fromarray(arr) |
| except Exception: |
| pass |
| if "path" in obj and obj["path"] and os.path.exists(str(obj["path"])): |
| try: |
| img = Image.open(obj["path"]) |
| img.load() |
| return img |
| except Exception: |
| pass |
| return None |
|
|
|
|
| def save_image(img: Image.Image, save_path: Path): |
| save_path.parent.mkdir(parents=True, exist_ok=True) |
| if img.mode not in ["1", "L", "LA", "P", "RGB", "RGBA", "I", "I;16"]: |
| img = img.convert("RGB") |
| img.save(save_path) |
|
|
|
|
| def unique_save_path(folder: Path, stem: str, ext: str) -> Path: |
| p = folder / f"{stem}{ext}" |
| k = 1 |
| while p.exists(): |
| p = folder / f"{stem}_{k}{ext}" |
| k += 1 |
| return p |
|
|
|
|
| def resolve_class_name(raw_label) -> str | None: |
| try: |
| tissue_id = int(float(str(raw_label).strip())) |
| except (ValueError, TypeError): |
| return None |
| return TISSUE_ID_TO_NAME.get(tissue_id, None) |
|
|
|
|
| def process_parquet(parquet_path: Path, fold_num: int, split_mapping: dict[str, str]) -> dict: |
| print(f"\n[INFO] Reading fold{fold_num}: {parquet_path.name}") |
| df = pd.read_parquet(parquet_path) |
| print(f"[INFO] Rows: {len(df)} | Columns: {list(df.columns)}") |
|
|
| class_counts: dict[str, dict[str, int]] = {} |
| saved = 0 |
| failed = 0 |
| no_split = 0 |
|
|
| for local_idx, row in enumerate(df.itertuples(index=False)): |
| stem = f"fold{fold_num}_{local_idx:08d}" |
|
|
| raw_label = getattr(row, CLASS_LABEL_COLUMN) |
| if is_null_like(raw_label): |
| print(f" [WARN] {stem} has empty class label, skipped") |
| failed += 1 |
| continue |
|
|
| class_name = resolve_class_name(raw_label) |
| if class_name is None: |
| print(f" [WARN] {stem} label '{raw_label}' is not in mapping table, skipped") |
| failed += 1 |
| continue |
|
|
| relative_path = f"{class_name}/{stem}{IMAGE_EXT}" |
| split = split_mapping.get(relative_path, None) |
|
|
| if split is None: |
| alt_stem = f"fold{fold_num}_{local_idx}" |
| alt_path = f"{class_name}/{alt_stem}{IMAGE_EXT}" |
| split = split_mapping.get(alt_path, None) |
|
|
| if split is None: |
| print(f" [WARN] {relative_path} not found in split CSV, skipped") |
| no_split += 1 |
| failed += 1 |
| continue |
|
|
| if split == "train": |
| output_root = OUTPUT_TRAIN_ROOT |
| elif split == "test": |
| output_root = OUTPUT_TEST_ROOT |
| else: |
| print(f" [WARN] {relative_path} has unknown split '{split}', skipped") |
| failed += 1 |
| continue |
|
|
| img = try_decode_image(getattr(row, IMAGE_COLUMN)) |
| if img is None: |
| print(f" [WARN] {stem} failed to decode image, skipped") |
| failed += 1 |
| continue |
|
|
| out_folder = output_root / class_name |
| save_path = unique_save_path(out_folder, stem, IMAGE_EXT) |
|
|
| try: |
| save_image(img, save_path) |
| if class_name not in class_counts: |
| class_counts[class_name] = {"train": 0, "test": 0} |
| class_counts[class_name][split] += 1 |
| saved += 1 |
| except Exception as e: |
| print(f" [ERROR] Failed to save ({save_path}): {e}") |
| failed += 1 |
|
|
| done = saved + failed |
| if done % 500 == 0: |
| print(f" ... Processed {done}/{len(df)} (saved {saved} / failed {failed})") |
|
|
| print(f"[INFO] fold{fold_num} finished: saved {saved} / failed {failed} " |
| f"(no_split_info: {no_split})") |
| return class_counts |
|
|
|
|
| def print_summary(total_class_counts: dict): |
| print("\n" + "=" * 70) |
| print(f"[DONE] Train directory: {OUTPUT_TRAIN_ROOT}") |
| print(f"[DONE] Test directory: {OUTPUT_TEST_ROOT}") |
| print(f"[DONE] Class statistics (total {len(total_class_counts)} classes):") |
| print(f" {'Class':<20s} {'Train':>8s} {'Test':>8s} {'Total':>8s}") |
| print(f" {'-'*20} {'-'*8} {'-'*8} {'-'*8}") |
|
|
| total_train = 0 |
| total_test = 0 |
| for cls in CLASS_ORDER: |
| counts = total_class_counts.get(cls, {"train": 0, "test": 0}) |
| train_n = counts.get("train", 0) |
| test_n = counts.get("test", 0) |
| total_train += train_n |
| total_test += test_n |
| print(f" {cls:<20s} {train_n:>8d} {test_n:>8d} {train_n + test_n:>8d}") |
|
|
| print(f" {'-'*20} {'-'*8} {'-'*8} {'-'*8}") |
| print(f" {'TOTAL':<20s} {total_train:>8d} {total_test:>8d} {total_train + total_test:>8d}") |
| print("=" * 70) |
|
|
|
|
| def main(): |
| for root in [OUTPUT_TRAIN_ROOT, OUTPUT_TEST_ROOT]: |
| root.mkdir(parents=True, exist_ok=True) |
| for cls in CLASS_ORDER: |
| (root / cls).mkdir(parents=True, exist_ok=True) |
|
|
| split_mapping = load_split_mapping(SPLIT_CSV_PATH) |
|
|
| total_class_counts: dict[str, dict[str, int]] = {} |
|
|
| for fold_num, parquet_path_str in enumerate(INPUT_PARQUET_PATHS, start=1): |
| parquet_path = Path(parquet_path_str) |
| if not parquet_path.exists(): |
| raise FileNotFoundError(f"Parquet file not found: {parquet_path}") |
|
|
| class_counts = process_parquet(parquet_path, fold_num, split_mapping) |
| for cls, counts in class_counts.items(): |
| if cls not in total_class_counts: |
| total_class_counts[cls] = {"train": 0, "test": 0} |
| for split_key in ["train", "test"]: |
| total_class_counts[cls][split_key] += counts.get(split_key, 0) |
|
|
| print_summary(total_class_counts) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|