Basically, here are 2 main process for the OCT(A) volume data First, reconstruct 3D raw volume data by stacking .bmp images Second, transform and resize the reconstructed raw data(linear registraion, crop&resize and resize nifti). The reason of these process is really clear. In the process of scanning, it is dificult to prevent all the movement from the patients. Especially the OCT data is a retina image, it is almost nonsense that we can control our pupil movement. And the data consists of left and right eyes, partially, it has structural differences. To deal with these points, we propose below data pre-processing steps.

• bmp2nii.py

• Convert Bmp files to Nifti files and set the 3 types of options.

  1. bmp_dir - load directory

     • consists of bmp files of OCT or OCTA image.
     • directory structure should be like below
     • bmp directory name should include OCT, OCTA.

  2. nii_dir - save directory

  3. FOV (Field Of View for scanning)

     • physical resolution = FOV size * scale
     • orgin, spacing will be setted by this FOV and scale factor.
     • This figure needs to set the header information of the Nifti.
     • source : IEEE-OCTA500

Transformation consists of several setps and overall process is shown in below Figure.

• Retina Layer segmentation

  • This part had been implemented by OCTExplorer3.8.0

• registration.py

  • In this experiment, to retain the volumetric data and try not to give a damage to the original data, only linear registration (translation, affine transformaion) had been leveraged.
  • However, we implemented the non-linear transformation either, you can use it if you want.
  • The library that we have used for registration is SimpleITK and you can install easily like below :

    • pip install SimpleITK
      pip install SimpleITK-SimpleElastix

• resize_nifti.py

  • The given data had been collected with different FOV, it has different resolution. To utilize the total data(500), it is needed to implement Volume resizing.
  • For FOV66(subject - 10001 ~ 10300), center-cropping has been used to match the FOV33 but as its size still 200x200x640, considering the pooling step for train-phase, we resize it to 192x192x640.
  • For FOV33(subject - 10301 ~ 10500), only the resizing had been adapted.
  • crop_resize.py is for the 2D. As 2D Enface has also different FOV, resizing is also needed either.

• check_volume.py

  • This script is for checking the height of all the volumes by considering non-zero value for the z-axis.

    • Procedure

      1. Check one-by-one and save the each first and last non-zero values.
      2. Before finish checking, save these z-indices for the .csv file.
      3. Choose the proper height based on the .csv file.
      4. Re-implement this module with the selected height.
      5. Then volume will be cropped from the middle point of the non-zero z indices with the selected height automatically.