dataset convert parquet file to nii.gz file

  # 配置路径规则
  dir_config = {
      "default": {"data": "data", "metadata": "metadata"},
      "unilateral": {"data": "data_unilateral", "metadata": "metadata_unilateral"},
  }

  # 创建输出路径
  output_root.mkdir(parents=True, exist_ok=True)

  # ------------------ 处理所有 .parquet 文件 ------------------
  parquet_files = sorted(parquet_dir.glob("*.parquet"))
  all_metadata = []

  for file_path in tqdm(parquet_files, desc="Processing .parquet files"):
      print(f"\n📂 正在处理: {file_path.name}")
      df = pd.read_parquet(file_path, engine="pyarrow")

      for idx, row in tqdm(df.iterrows(), total=len(df), desc="Saving images"):
          uid = row["UID"]
          institution = row["Institution"]
          # 获取图像名称
          img_names = [col.split("Image_")[1] for col in row.index if col.startswith("Image_")]
          # 输出路径
          path_folder = output_root / institution / dir_config[config]["data"] / uid
          path_folder.mkdir(parents=True, exist_ok=True)
          item_metadata = {
              "UID": uid,
              "Institution": institution,
              "Split": row.get("Split", "unknown"),
              "Fold": row.get("Fold", -1),
          }
          for img_name in img_names:
              img_key = f"Image_{img_name}"
              affine_key = f"Affine_{img_name}"
              img_data = row[img_key]
              img_affine = row[affine_key]
              if img_data is None or affine_key not in row:
                  continue
              img_np = deep_unpack(img_data)
              affine_np = deep_unpack(img_affine)
              print(f"🔍 正在处理 UID: {uid}, 图像: {img_name}, 数据类型: {type(img_data)}")
              print(f"📏 图像 shape: {img_np.shape}, dtype: {img_np.dtype}")
              print(f"📏 affine shape: {affine_np.shape}, dtype: {img_affine.dtype}")
              print(f"🔬 图像样本类型: {type(img_np.flat[0])}")  # 检查元素
              if img_np.dtype == object:
                  print("⚠️ 图像是 object 类型数组，内部结构可能不规则")
              try:
                  img_np = np.array(img_np, dtype=np.int16)
                  affine_np = np.array(affine_np, dtype=np.float64)
                  if img_np.ndim == 3:
                      img_np = np.expand_dims(img_np, axis=0)
                      img_tensor = tio.ScalarImage(tensor=img_np, affine=affine_np)
                      img_tensor.save(path_folder / f"{img_name}.nii.gz")
                  else:
                      print(f"❗跳过 UID: {uid} 图像 {img_name}，维度不正确: {img_np.shape}")
              except Exception as e:
                  print(f"❌ 图像转换失败 UID: {uid}, 错误: {e}")
          all_metadata.append(item_metadata)
  # ------------------ 保存标签 ------------------
  df_all = pd.DataFrame(all_metadata)
  for institution in df_all["Institution"].unique():
      df_inst = df_all[df_all["Institution"] == institution]
      path_metadata = output_root / institution / dir_config[config]["metadata"]
      path_metadata.mkdir(parents=True, exist_ok=True)
      df_anno = df_inst.drop(columns=["Institution", "Split", "Fold"])
      df_anno.to_csv(path_metadata / "annotation.csv", index=False)
      df_split = df_inst[["UID", "Split", "Fold"]]
      df_split.to_csv(path_metadata / "split.csv", index=False)
  print("✅ 所有图像与标签保存完成！")