import os import csv import json import shutil import re import numpy as np import pandas as pd from pathlib import Path from PIL import Image, UnidentifiedImageError from typing import List, Dict, Any, Set, Tuple DOWNLOAD_ROOT = "./Skin_Image" EXTRACT_ROOT = "./Skin_Image" OUTPUT_IMAGE_DIR = "./Skin_Image/images" IMAGE_SIZE = 512 MIN_WIDTH = 28 MIN_HEIGHT = 28 DATASET_CONFIGS = [ { "bench_class": "Isic2018MSBench", "module": "medsegbench", "npz_stem": "isic2018_512", "dataset_name": "isic2018", }, { "bench_class": "UWSkinCancerMSBench", "module": "medsegbench", "npz_stem": "uwaterlooskincancer_512", "dataset_name": "uwaterlooskincancer", }, ] def to_uint8(arr): arr = np.asarray(arr) if arr.dtype == np.uint8: return arr arr = arr.astype(np.float32) if arr.size > 0 and arr.min() >= 0 and arr.max() <= 1.0: arr = arr * 255.0 return np.clip(arr, 0, 255).astype(np.uint8) def save_image(arr, save_path): arr = to_uint8(arr) if arr.ndim == 2: img = Image.fromarray(arr, mode="L") elif arr.ndim == 3: if arr.shape[2] == 1: img = Image.fromarray(arr[:, :, 0], mode="L") elif arr.shape[2] == 3: img = Image.fromarray(arr, mode="RGB") elif arr.shape[2] == 4: img = Image.fromarray(arr, mode="RGBA") else: raise ValueError(f"Unsupported image shape: {arr.shape}") else: raise ValueError(f"Unsupported image shape: {arr.shape}") img.save(save_path) def is_binary_mask(arr): unique_vals = set(np.unique(arr).tolist()) return unique_vals.issubset({0, 1}) def key_to_subdir(key): parts = key.split("_") if len(parts) >= 2 and parts[0] in {"train", "val", "valid", "validation", "test"}: split = parts[0] rest = "_".join(parts[1:]) return Path(split) / rest return Path(key) def mask_to_bbox(mask_arr): if mask_arr.ndim == 3: mask_arr = mask_arr[:, :, 0] ys, xs = np.where(mask_arr > 0) if len(xs) == 0: return None return int(xs.min()), int(ys.min()), int(xs.max()) + 1, int(ys.max()) + 1 def step1_download_datasets(): print("\n" + "=" * 70) print("STEP 1: Download datasets") print("=" * 70) import importlib for cfg in DATASET_CONFIGS: mod = importlib.import_module(cfg["module"]) bench_cls = getattr(mod, cfg["bench_class"]) print(f"\n[Download] {cfg['bench_class']} -> {DOWNLOAD_ROOT}") _ = bench_cls( root=DOWNLOAD_ROOT, split="train", download=True, size=IMAGE_SIZE, ) print(f"[Done] {cfg['bench_class']} download/verification complete") def step2_extract_npz(cfg) -> Path: npz_path = Path(DOWNLOAD_ROOT) / f"{cfg['npz_stem']}.npz" if not npz_path.exists(): raise FileNotFoundError(f"NPZ file not found: {npz_path}") out_root = Path(EXTRACT_ROOT) / cfg["npz_stem"] out_root.mkdir(parents=True, exist_ok=True) print(f"\n[Extract] {npz_path}") data = np.load(str(npz_path), allow_pickle=True) print(" Keys:") for key in data.files: arr = data[key] print(f" - {key}: shape={arr.shape}, dtype={arr.dtype}") item_count = 0 for key in data.files: arr = np.asarray(data[key]) subdir = key_to_subdir(key) target_dir = out_root / subdir target_dir.mkdir(parents=True, exist_ok=True) if arr.ndim == 3 and arr.shape[-1] not in (1, 3, 4): for i in range(arr.shape[0]): single = arr[i] if is_binary_mask(single): np.save(target_dir / f"{i:05d}.npy", single) else: save_image(single, target_dir / f"{i:05d}.png") item_count += 1 elif arr.ndim == 4 and arr.shape[-1] in (1, 3, 4): for i in range(arr.shape[0]): save_image(arr[i], target_dir / f"{i:05d}.png") item_count += 1 elif arr.ndim == 2 or (arr.ndim == 3 and arr.shape[-1] in (1, 3, 4)): if is_binary_mask(arr): np.save(target_dir / "0.npy", arr) else: save_image(arr, target_dir / "0.png") item_count += 1 else: np.save(target_dir / f"{key}.npy", arr) item_count += 1 print(f" Extraction complete: {item_count} items -> {out_root}") return out_root def step3_crop(cfg, extract_root: Path) -> List[Tuple[str, str]]: print(f"\n[Crop] Dataset: {cfg['dataset_name']}") os.makedirs(OUTPUT_IMAGE_DIR, exist_ok=True) dataset_name = cfg["dataset_name"] saved_pairs = [] for split_dir in sorted(extract_root.iterdir()): if not split_dir.is_dir(): continue split_name = split_dir.name if split_name.startswith("."): continue images_dir = split_dir / "images" labels_dir = None for label_candidate in ["labels", "label", "label_C1", "masks", "mask"]: candidate = split_dir / label_candidate if candidate.is_dir(): labels_dir = candidate break if not images_dir.is_dir(): print(f" [Skip] No images directory: {split_dir}") continue if labels_dir is None: print(f" [Skip] No labels directory: {split_dir}") continue image_files = sorted([ f for f in images_dir.iterdir() if f.suffix.lower() in {".png", ".jpg", ".jpeg", ".bmp", ".tif", ".tiff", ".npy"} ]) print(f" [{split_name}] Image count: {len(image_files)}, labels dir: {labels_dir.name}") for img_file in image_files: idx_str = img_file.stem if img_file.suffix == ".npy": img_arr = np.load(str(img_file)) img_arr = to_uint8(img_arr) if img_arr.ndim == 2: pil_img = Image.fromarray(img_arr, mode="L").convert("RGB") elif img_arr.ndim == 3 and img_arr.shape[2] == 1: pil_img = Image.fromarray(img_arr[:, :, 0], mode="L").convert("RGB") elif img_arr.ndim == 3: pil_img = Image.fromarray(img_arr, mode="RGB") else: continue else: try: pil_img = Image.open(str(img_file)).convert("RGB") except Exception as e: print(f" [WARN] Failed to open image: {img_file}, {e}") continue mask_arr = None for ext in [".npy", ".png", ".jpg", ".bmp", ".tif"]: mask_file = labels_dir / f"{idx_str}{ext}" if mask_file.exists(): if ext == ".npy": mask_arr = np.load(str(mask_file)) else: mask_arr = np.array(Image.open(str(mask_file)).convert("L")) break if mask_arr is None: continue bbox = mask_to_bbox(mask_arr) if bbox is None: continue x_min, y_min, x_max, y_max = bbox w, h = pil_img.size x_min = max(0, min(x_min, w - 1)) y_min = max(0, min(y_min, h - 1)) x_max = max(0, min(x_max, w)) y_max = max(0, min(y_max, h)) if x_max <= x_min or y_max <= y_min: continue new_image_name = f"{split_name}_{idx_str}_{dataset_name}.png" crop_image_name = f"{split_name}_{idx_str}_{dataset_name}_crop.png" new_image_path = os.path.join(OUTPUT_IMAGE_DIR, new_image_name) crop_image_path = os.path.join(OUTPUT_IMAGE_DIR, crop_image_name) try: pil_img.save(new_image_path, format="PNG") except Exception as e: print(f" [WARN] Failed to save original: {new_image_path}, {e}") continue crop = pil_img.crop((x_min, y_min, x_max, y_max)) try: crop.save(crop_image_path, format="PNG") except Exception as e: print(f" [WARN] Failed to save crop: {crop_image_path}, {e}") continue saved_pairs.append((new_image_path, crop_image_path)) print(f" Saved {len(saved_pairs)} image pairs") return saved_pairs def step4_filter_small_images(saved_pairs: List[Tuple[str, str]]): print(f"\n[Filter] Checking {len(saved_pairs)} image pairs") deleted_count = 0 for original_path, crop_path in saved_pairs: remove = False if os.path.isfile(crop_path): try: with Image.open(crop_path) as img: w, h = img.size if w < MIN_WIDTH or h < MIN_HEIGHT: remove = True print(f" [Small] {os.path.basename(crop_path)}: {w}x{h}") except (UnidentifiedImageError, OSError) as e: remove = True print(f" [Broken] {crop_path}: {e}") else: remove = True if remove: deleted_count += 1 if os.path.isfile(crop_path): os.remove(crop_path) if os.path.isfile(original_path): os.remove(original_path) kept = len(saved_pairs) - deleted_count print(f" Total: {len(saved_pairs)} | Deleted: {deleted_count} | Kept: {kept}") def main(): step1_download_datasets() for cfg in DATASET_CONFIGS: print("\n" + "=" * 70) print(f"Processing dataset: {cfg['dataset_name']}") print("=" * 70) extract_root = step2_extract_npz(cfg) saved_pairs = step3_crop(cfg, extract_root) step4_filter_small_images(saved_pairs) print("\n" + "=" * 70) print("All processing complete") print("=" * 70) print(f"Output image directory: {OUTPUT_IMAGE_DIR}") if os.path.isdir(OUTPUT_IMAGE_DIR): all_images = [f for f in os.listdir(OUTPUT_IMAGE_DIR) if f.endswith(".png")] originals = [f for f in all_images if not f.endswith("_crop.png")] crops = [f for f in all_images if f.endswith("_crop.png")] print(f"Final image count: {len(all_images)} (originals: {len(originals)}, crops: {len(crops)})") if __name__ == "__main__": main()