使用np.load加载DICOM图像失败及维度异常问题求助

问题1：IMAGE_SPATIAL_DIMS输出为3而非预期三维形状 #

运行以下代码时：

import numpy as np
...
file_name = "/content/drive/MyDrive/.../1-01.dcm"
IMAGE_SPATIAL_DIMS = np.load(file_name)[...,0].shape

预期得到类似(128, 128, 16)的三维形状，实际输出结果为3。

问题2：加载预处理图像时的Pickle错误 #

• 初始报错：ValueError: Cannot load file containing pickled data when allow_pickle=False
• 设置allow_pickle=True后仍报错：UnpicklingError: Failed to interpret file '1-01.dcm' as a pickle

预处理代码 #

import os
import sys
import shutil
import numpy as np
import pandas as pd
import pydicom as dicom
from os.path import join
import SimpleITK as sitk
from spire.xls import *
from spire.xls.common import *
import openpyxl as op
import time

### Read TrainFold0 Excel Sheet
CODE_BASE = os.path.abspath('/content/drive/MyDrive/MasterThesis')
TRAIN_XLSX = '/content/drive/MyDrive/MasterThesis/.../Datasets/TrainFold0.xlsx'

### Make DataFrame of 3D Data
dataFrame = pd.read_excel(TRAIN_XLSX)

### Create First Row of DataFrame3D
dataFrame3D = pd.DataFrame()

### Make 3D Train Data Directory
Data3D_path = '/content/drive/MyDrive/MasterThesis/.../Datasets/3DPROSTATEx'
if not(os.path.exists(Data3D_path)):
  os.mkdir(Data3D_path)
### Load an TrainFold Excel File
wb = op.load_workbook(filename=TRAIN_XLSX)
ws = wb.active

### Read DataFrame of TrainFold Excel File
dataFrame = pd.read_excel(TRAIN_XLSX)
col_imagePath = dataFrame['image_path']

### Get the Number of Rows
row_num = len(col_imagePath)

### Read the image_path of Every Row
last_indx = 0
for row_count in range(row_num):
  print('row_count: '+str(row_count))
  path = col_imagePath[row_count]
  path_3D =  "/".join(path.split('/')[2:-1])
  #print(path)
  #print(path_3D)
  ### Make Directory for Image3D
  for i in range(1, len(path_3D.split('/'))+1):
    if not(os.path.exists(Data3D_path + '/' + "/".join(path_3D.split('/')[:i]))):
      os.mkdir(Data3D_path + '/' + "/".join(path_3D.split('/')[:i]))

  ### Make a New Name for 3D Image
  #e.g : path_3D = 'ProstateX-0030/10-30-2011-NA-MC prostaat kliniek detectie-mc MCPROSKL30-54485/3.000000-t2tsesag-6352'
  part1  = path_3D.split('/')[0]                                 #e.g: ProstateX-0030
  part11 = part1.split('-')[1]                                   #e.g: 0030

  part2  = path.split('/')[-1]                                   #e.g: 3.000000-t2tsesag-6352
  part21 = part2.split('-')[0]                                   #e.g: 3.000000
  part22 = part21.split('.')[0]                                  #e.g: 3

  image3D_Name = part11+ '-' + str(part22).rjust(2, '0')+'.dcm'  #e.g: 0030-03.dcm
  #time.sleep(5)
  ### If Related 3D Image Doesn't Exist Create it
  if (os.path.exists(Data3D_path +'/'+ path_3D+'/'+image3D_Name)):
    print(str(image3D_Name)+'Image Does not Exist')
    ### Get Images in Directory
    os.chdir('/content/drive/MyDrive/MasterThesis/.../Datasets'+path)
    entries = os.listdir('/content/drive/MyDrive/MasterThesis/.../Datasets'+path)

    image3D = []
    for entry in entries:
      ### Check if Entry doesn't Image ,pass away this loop and go to the next entry
      if not(os.path.isfile(entry)):
        continue

      ### Read DICOM Image as a Slice
      #new_slice = dicom.dcmread(entry).pixel_array

      ### Read DICOM Image with numpy.load
      new_slice = np.load(entry)
      ### Expand Dimension of Slice Image from [Height, Width] to [Height, Width, 1]
      new_slice3D = np.expand_dims(new_slice, axis=2)
      if len(image3D)!=0 :
        ### Append New Slice to 3D Image
        image3D = np.append(image3D, new_slice3D, axis=2)
      else:
        ### Create Frist Slice
        image3D = new_slice3D

    ### Save Image3D in Related Directory
    os.chdir(Data3D_path +'/'+ path_3D)
    iMAGE_TRAIN_3D = sitk.GetImageFromArray(image3D)
    sitk.WriteImage(iMAGE_TRAIN_3D, Data3D_path +'/'+ path_3D +'/'+ image3D_Name)

  ### Export Path of Image3D in DataFrame3D
  new_row = dataFrame.iloc[[row_count]].copy()

  ### Concatenate New Row to dataFrame3D
  dataFrame3D = pd.concat([dataFrame3D, dataFrame.iloc[[row_count]].copy()] , axis=0, ignore_index=True)

  ### Change Old Path to New Path
  dataFrame3D.loc[dataFrame3D.index[-1],'image_path'] = Data3D_path+'/'+ path_3D+'/'+ image3D_Name

  ### Convert DataFrame3D to TrainFold3D Excel File
  dataFrame3D.to_csv(join(TRAIN_XLSX_PREFIX+'3D.csv'))

问题分析与解决 #

核心错误根源 #

你全程误用了np.load()读取DICOM文件：np.load()仅支持读取numpy专属的.npy/.npz格式文件，而DICOM是医学影像专用格式，两者完全不兼容，这是所有问题的起点。

1. 解决shape输出为3的问题 #

改用DICOM专用库读取文件，再转换为numpy数组获取形状：

方法1：用SimpleITK读取（匹配你预处理时的保存方式）

import SimpleITK as sitk
file_name = "/content/drive/MyDrive/.../1-01.dcm"
# 读取DICOM文件
sitk_image = sitk.ReadImage(file_name)
# 转换为numpy数组（注意顺序为(z, y, x)）
image_array = sitk.GetArrayFromImage(sitk_image)
# 若需要(y, x, z)的顺序，可转置
IMAGE_SPATIAL_DIMS = image_array.transpose(1, 2, 0).shape

方法2：用pydicom读取

import pydicom
file_name = "/content/drive/MyDrive/.../1-01.dcm"
ds = pydicom.dcmread(file_name)
# 多帧DICOM的pixel_array为(z, y, x)形状
IMAGE_SPATIAL_DIMS = ds.pixel_array.shape

2. 解决Pickle错误问题 #

直接弃用np.load()读取DICOM文件，改用上述两种DICOM专用读取方式即可，无需调整allow_pickle参数。

预处理代码的关键修正 #

1. 修正切片读取逻辑：把错误的new_slice = np.load(entry)替换为正确的DICOM读取代码：

   # 替换原错误代码
   ds = dicom.dcmread(entry)
   new_slice = ds.pixel_array
   

2. 修正文件存在判断逻辑：原代码中if (os.path.exists(...)):逻辑颠倒，应改为：

   if not os.path.exists(Data3D_path +'/'+ path_3D+'/'+image3D_Name):
   

3. 避免工作目录切换：用绝对路径拼接代替os.chdir，防止路径混乱：

   slice_dir = '/content/drive/MyDrive/MasterThesis/.../Datasets'+path
   entries = os.listdir(slice_dir)
   for entry in entries:
       entry_path = os.path.join(slice_dir, entry)
       if not os.path.isfile(entry_path):
           continue
       # 后续读取操作使用entry_path
   

4. 优化DataFrame保存：把dataFrame3D.to_csv()移到循环结束后，避免重复写入浪费资源。

内容的提问来源于stack exchange，提问作者zahra mrtz