OCTIP: OCT Image Preprocessing

This package was developed for the analysis of Spectralis OCT images (in XML+PNG or XML+BMP format). It essentially contains image preprocessing code, optionally relying on a retina segmentation step.

Installation

Install TensorFlow using PIP, preferably using VirtualEnv.

This package was tested successfully with Tensorflow version 2.1.0 and Python version 3.6.8.

Then, install OCTIP in developer mode (option -e):

pip install -e octip

Usage Examples

The latter version of OCTIP comes with two scripts:

• octip-spectralis-2-nifti.py converts Spectralis OCT dataset in compressed NifTI format,
• octip-dataset-split.py splits the converted dataset into train, validation and test subsets.

Let OCT_Images denote a directory containing OCT data with the following structure:

• Each subdirectory of OCT_Images is associated with one patient.
• Each patient's directory contains any hierarchy of subdirectories. Any directory in this hierarchy that contains at least one .xml file and either a .bmp or .png file is considered an OCT volume.

Follow the instructions below to:

1. convert the dataset, optionally using retina segmentation.
2. split the dataset into train/validation/test subsets.

Dataset Conversion Without Retina Segmentation

The following bash command simply converts each eye exam to a volume of 32x224x224 voxels (or more generally depthxheightxwidth voxels) by:

• selecting 32 B-scans (or more generally depth B-scans) per volume,
• resizing each selected B-scan to 224x224 pixels (or more generally heightxwidth pixels),
• saving the volume as volumes/patient<x>-eye-exam<y>.nii.gz, where <x> is a unique patient identifier and <y> identifies one exam of one eye of the patient.

octip-spectralis-2-nifti.py --depth 32 --height 224 --width 224 --input_dirs ../OCT_Images

There can be more than one input directory.

Note that the 32 B-scans are selected as follows:

• if the original volume has less then 32 B-scans, then all B-scans are selected and empty (black) B-scans are added before and after these B-scans to obtain a total of 32 B-scans.
• if the original volume has 32 B-scans or more, then 32 B-scans are selected uniformly.

Dataset Conversion With Retina Segmentation

The following bash command converts each eye exam to a volume of 32x192x224 voxels (or more generally depthxheightxwidth voxels) using retina segmentation:

CUDA_VISIBLE_DEVICES=0 octip-spectralis-2-nifti.py -r octip_models --depth 32 --height 192 --width 224 \
    --input_dirs ../OCT_Images

Request directory octip_models containing the following two model files (gwenole.quellec@inserm.fr):

• FPN_efficientnetb6_384x384.hdf5,
• FPN_efficientnetb7_320x320.hdf5.

The 32 B-scans (or more generally depth B-scans) are selected similarly to the previous example and the ground truth file is identical. Differences with the previous example are that:

• the retina is segmented in each selected B-scans and only the first 192 pixels (more generally height pixels) below the retinal top surface are selected,
• optionally (if the --normalize_intensities option is set), intensity in the selected region is normalized,
• the selected region is resized to 192x224 pixels (or more generally heightxwidth pixels).

Based on initial experiments, the --normalize_intensities option seems relevant for progression measurement between two exams. However, it does not seem relevant for single exam classification.

Viewing the Converted NifTI Volumes

ImageJ's nifti plugin can be used to view the compressed NifTI files. After installing the plugin, a volume can be imported in ImageJ using File -> Import -> NifTI/Analyze. For convenience, volumes are also stored as directories of PNG images in the resized_image directory.

Splitting the Converted Dataset into Training, Validation and Test Subsets

Run the following command to split a ground truth file ground-truth.csv associated with such a dataset into training, validation and test subsets:

octip-dataset-split.py -t ground-truth.csv --ratios 0.8 0.1 0.1 --patient_col patient \
    --labels DRUSEN AMD DME

This will assign 80% of the patients to the training subset and 10% of them to the validation and test subsets. More generally, the train:validation:test ratios are given by ratios[0]:ratios[1]:ratios[2]. This script ensures that:

• the listed labels (e.g., DRUSEN, AMD, DME) are distributed across the three subsets as evenly as possible,
• all exams from the same patient (according to column patient of ground-truth.csv) are assigned to the same subset.

Note that this script only splits the ground truth file into subsets (train.csv, validation.csv and test.csv). The volumes directory does not need to be split. Simply use symbolic links if need be (e.g. for compatibility with my libraries):

ln -s volumes train
ln -s volumes validation
ln -s volumes test