Segmentation disagreement between dicom2nifti and Monai when converting DICOM to NIFTI format

[staydelight]

Environment

• OS:
  Linux

• Python version:
  3.8.18

• Monai version:
  1.3.0

• dicom2nifti version:
  2.4.8

• Dataset:

  https://wiki.cancerimagingarchive.net/display/Public/PROSTATE-DIAGNOSIS

  Test data path:

   manifest-gJIZVVFt6412408718812805737\
       PROSTATE-DIAGNOSIS\
          ProstateDx-01-0001\
              09-23-2008-NA-MRI PROSTATE WITH AND WITHOUT CONTRAST-00237\
                 501.000000-T2WTSEAX-40778
  

Steps to reproduce the disagreement

• Process dicom image

  No preprocessing of the dataset is required, the input is a DICOM file with the following command:

   python -m monai.bundle run --config_file configs/inference_dicom.json
  

  The config file is referred from 'Project-MONAI/modelzoo/models/prostate_mri_anatomy/configs/inference.json' with the following changes to modify its path:

    "imports": [
        "$import pandas as pd",
+       "$import glob",
        "$import os"
    ],
-   "bundle_root": "/workspace/data/prostate_mri_anatomy",
+   "bundle_root": ".",
-   "output_dir": "$@bundle_root + '/eval'",
+   "output_dir": "$@bundle_root + '/eval_monai'",
-   "dataset_dir": "/workspace/data/prostate158/prostate158_test/",
+   "dataset_dir": "501.T2WTSEAX",
-    "datalist": "$list(@dataset_dir + pd.read_csv(@dataset_dir + 'test.csv').t2)",
+   "datalist": "$list(sorted(glob.glob(@dataset_dir)))",

• Process nifti image

  Step 1: Convert dicom to nifti with the following command:

  dicom2nifti 501.000000-T2WTSEAX-40778/ dicom2nifti.501
  

  Step 2: Get segmentation results from Monai model:

  python -m monai.bundle run --config_file configs/inference_nifti.json
  

  The config file is referred from the same source as Monai with the following changes to modify its path:

    "imports": [
        "$import pandas as pd",
+       "$import glob",
        "$import os"
    ],
-   "bundle_root": "/workspace/data/prostate_mri_anatomy",
+   "bundle_root": ".",
-   "output_dir": "$@bundle_root + '/eval'",
+   "output_dir": "$@bundle_root + '/eval_d2n'",
-   "dataset_dir": "/workspace/data/prostate158/prostate158_test/",
+   "dataset_dir": "dicom2nifti.501",
-   "datalist": "$list(@dataset_dir + pd.read_csv(@dataset_dir + 'test.csv').t2)",
+   "datalist": "$list(sorted(glob.glob(@dataset_dir+'/*.nii.gz')))",

Expected result

The segmentation results should be the same.

My investigation

As you can see from the pictures below , two segmentation results are different:

Process dicom image	Process nifti image

After digging into the code a little bit, I find that the affine matrix should play the important role in making this difference.

Finding the affine matrix

• Process dicom image

  The affine matrix from 'monai/data/image_reader.py' is defined with:

  direction = np.asarray(direction)
  sr = min(max(direction.shape[0], 1), 3)
  affine: np.ndarray = np.eye(sr + 1)
  affine[:sr, :sr] = direction[:sr, :sr] @ np.diag(spacing[:sr])
  affine[:sr, -1] = origin[:sr]
  if lps_to_ras:
      affine = orientation_ras_lps(affine)
    
  self.logger.debug(f"Affine matrix:\n{affine}")
  
  return affine
  

  From the input that we given, we can find affine matrix with the following:

  

• Process nifti image

  The affine matrix from 'dicom2nifti/common.py' is defined with:

  affine = numpy.array(
          [[-image_orient1[0] * delta_c, -image_orient2[0] * delta_r, -step[0], -image_pos[0]],
          [-image_orient1[1] * delta_c, -image_orient2[1] * delta_r, -step[1], -image_pos[1]],
          [image_orient1[2] * delta_c, image_orient2[2] * delta_r, step[2], image_pos[2]],
          [0, 0, 0, 1]]
  )
    
  logger.debug(f"Affine matrix:\n{affine}")
  
  return affine, numpy.linalg.norm(step)
  

  From the input that we given, we can find affine matrix with the following:

  

  The difference:

  Process dicom image	Process nifti image

Truncate the floating point

I added this line separately to both Monai and dicom2nifti:

affine = orientation_ras_lps(affine)

After I truncated these individual values to five decimal places, we can achieve the same affine matrix and the same segmentation results.

Process dicom image	Process nifti image

It is obvious that this element, with its precise floating-point value plays a role.
Given this result, I don't know which is correct, so I am sharing them here to seek feedback.