### 加载和处理 nii.gz 格式的医学影像数据集 在 PyTorch 中加载和处理 `.nii.gz` 文件通常涉及以下几个方面：读取文件、转换为张量以及将其集成到 `Dataset` 和 `DataLoader` 中以便于训练模型。 #### 使用 NiBabel 库读取 .nii.gz 数据 `.nii.gz` 是一种常见的医学图像存储格式，可以使用 Python 的第三方库 **NiBabel** 来读取这些文件。安装该库可以通过 pip 完成： ```bash pip install nibabel ``` 以下是通过 NiBabel 读取 `.nii.gz` 文件并将其转换为 NumPy 数组的代码示例[^1]： ```python import nibabel as nib import numpy as np def load_nifti_file(file_path): """ Load a NIfTI file and return its data as a NumPy array. Parameters: file_path (str): Path to the .nii or .nii.gz file. Returns: np.ndarray: The image data as a NumPy array. """ img = nib.load(file_path) # Load the NIfTI file using nibabel data = img.get_fdata() # Extract the raw data from the loaded object return data # Return it as a NumPy array ``` #### 将 NumPy 数组转换为 PyTorch 张量 为了使数据能够在 PyTorch 模型中使用，需要将 NumPy 数组转换为 PyTorch 张量。这一步非常简单，可以直接利用 `torch.from_numpy()` 方法完成[^2]： ```python import torch def convert_to_tensor(numpy_data): """ Convert a NumPy array into a PyTorch tensor. Parameters: numpy_data (np.ndarray): Input data in NumPy format. Returns: torch.Tensor: Data converted to PyTorch Tensor. """ tensor_data = torch.from_numpy(numpy_data).float() return tensor_data.unsqueeze(0) # Add channel dimension, e.g., shape becomes [1, D, H, W] ``` #### 创建自定义 Dataset 类 为了让数据能够被高效地批量加载，在 PyTorch 中推荐创建一个继承自 `torch.utils.data.Dataset` 的类来封装数据加载逻辑。下面是一个简单的实现例子[^3]： ```python from torch.utils.data import Dataset class MedicalImageDataset(Dataset): def __init__(self, file_paths, labels=None, transform=None): """ Initialize the dataset with paths of files and optional labels/transforms. Parameters: file_paths (list): List of strings representing file paths. labels (list, optional): Corresponding list of labels for each file path. transform (callable, optional): Optional transformation function applied on sample. """ self.file_paths = file_paths self.labels = labels if labels is not None else [None]*len(file_paths) self.transform = transform def __len__(self): """Return total number of samples.""" return len(self.file_paths) def __getitem__(self, idx): """Retrieve one item by index.""" file_path = self.file_paths[idx] label = self.labels[idx] # Load the MRI scan mri_scan = load_nifti_file(file_path) # Apply transformations if any are specified if self.transform: mri_scan = self.transform(mri_scan) # Convert to tensor mri_tensor = convert_to_tensor(mri_scan) if label is not None: return mri_tensor, label return mri_tensor ``` #### 集成 DataLoader 并进行批处理 最后，可以借助 `torch.utils.data.DataLoader` 对象轻松管理批次大小和其他参数设置。例如[^4]: ```python from torch.utils.data import DataLoader file_list = ['path/to/file1.nii.gz', 'path/to/file2.nii.gz'] # Example file list dataset = MedicalImageDataset(file_list) dataloader = DataLoader(dataset, batch_size=2, shuffle=True) for batch_idx, batch_data in enumerate(dataloader): print(f'Batch {batch_idx}: Shape={batch_data.shape}') ``` ---