Fractal features for proliferative diabetic retinopathy screening

This code corresponds to our Medical Physics paper with Karel van Keer, João Barbosa Breda, Hugo Luis Manterola, Matthew B. Blaschko and Alejandro Clausse, entitled "Proliferative Diabetic Retinopathy Characterization based on Fractal Features: Evaluation on a Publicly Available Data Set".

If you use this code in any publication, please include the following citation:

@article{orlando2017fractal,
	title={Proliferative Diabetic Retinopathy Characterization based on Fractal Features: Evaluation on a Publicly Available Data Set},
	author={Orlando, Jos\'e Ignacio and van Keer, Karel and Barbosa Breda, Jo\~ao Manterola, Hugo Luis and Blaschko, Matthew B. and Clausse, Alejandro},
	journal={Medical Physics},
	year={2017},
	publisher={Wiley}
}

If you use our code for red lesion detection, please cite the paper indicated in https://github.com/ignaciorlando/red-lesion-detection.

Abstract

Purpose: Diabetic retinopathy (DR) is one of the most widespread causes of preventable blindness in the world. The most dangerous stage of this condition is proliferative DR (PDR), in which the risk of vision loss is high and treatments are less effective. Fractal features of the retinal vasculature have been previously explored as potential biomarkers of DR, yet the current literature is inconclusive with respect to their correlation with PDR. In this study we experimentally assess their discrimination ability to recognize PDR cases.

Methods: A statistical analysis of the viability of using three reference fractal characterization schemes--namely box, information and correlation dimensions--to identify patients with PDR is presented. These descriptors are also evaluated as input features for training L1 and L2 regularized logistic regression classifiers, to estimate their performance.

Results: Our results on MESSIDOR, a public data set of 1200 fundus photographs, indicate that patients with PDR are more likely to exhibit a higher fractal dimension than healthy subjects or patients with mild levels of DR. Moreover, a supervised classifier trained with both fractal measurements and red lesion based features reports an area under the ROC curve of 0.9291 for PDR screening and 0.9574 for detecting patients with optic disc neovascularization.

Conclusions: The fractal dimension of the vasculature increases with the level of DR. Furthermore, PDR screening using multiscale fractal measurements is more feasible than using their derived fractal dimensions. Code and further resources are provided in this webpage.

Getting Started

These instructions will get you a copy of the project up and running on your local machine for development and testing purposes. .

Prerequisites

We make use of some external libraries in our code, namely:

• VLFeat: for efficient ROC curve computation. [link]
• Mark Schmidt implementation of regularized logistic regression. [link]
• Our red lesion detector. [link]

These two libraries are included by default in fundus-fractal-analysis/external.

It will also require to download MESSIDOR data set, which is available here. Vascular segmentations of this data set, obtained using this method, can be downloaded from here.

Installing

1. Create a fork of this repository, or clone it by doing:

$ git clone https://github.com/ignaciorlando/fundus-fractal-analysis

2. Open MATLAB (I used MATLAB R2015a) and move to fundus-fractal-analysis folder.
3. Run setup_fractal to compile all MEX files and to add the necessary folders to your MATLAB path.
4. Download MESSIDOR and segmentations.
5. Edit configuration_files/data_organization/config_organize_messidor_data to indicate where the downloaded MESSIDOR data set is located.
6. Run data_organization/script_organize_messidor_data to setup MESSIDOR in the correct form. This may take a while, as it will copy all the images to a new folder, preparing a .MAT files with the image labels, and it will generate field of view (FOV) masks for all the images in the data set and crop them around the FOV for faster processing.
7. Copy the segmentations that you downloaded to the new MESSIDOR folder.
8. Everything is ready to run!

Note: We tested this code on a Macbook PRO Retina 2015 with macOS Sierra, but in principle there would be no problem on using it within other architectures or operating systems.

Running experiments

Scripts for running experiments is in the experiments folder. All filenames are quite self-explanatory, but if you have any questions do not hesitate in opening issues in my github repository.

In general, all experiments will require to edit a configuration script. They are all in configuration_files.

Fractal feature extraction

Scripts for fractal feature extraction are in the feature-extraction folder, and their configuration files are in configuration_files/feature-extraction.

Two different scripts for extracting fractal features are available:

• Run script_extract_fractal_dimensions to get fractal dimensions from all the images in MESSIDOR.
• Run script_extract_fractal_features to get fractal measurements from all the images in MESSIDOR.

Red lesion feature extraction

• Run script_extract_red_lesion_features to segment red lesions and extract features from them (so far, the maximum probability).

Correlation analysis

• Run script_perform_correlation_analysis to estimate linear correlation between different fractal dimensions (Table 1 from our paper).

Boxplots with fractal dimensions distribution per DR grade

• Run script_boxplot_fractal_dimension_for_each_label to get boxplots depicting fractal dimensions distribution per DR grade (Figures 2 and 3 from our paper).

Boxplots with fractal dimensions distribution for proliferative DR screening

• Run script_boxplot_proliferative_dr_screening to get boxplots depicting fractal dimensions distribution for proliferative DR screening (Figures 4 and 5 from our paper).

Kolmogorov-Smirnof tests between DR grades

• Run script_ks_test_between_grades to perform Wilcoxon rank sum tests and estimate if differences in distributions between grades are statistically significant.

Kolmogorov-Smirnof tests for proliferative DR screening

• Run script_ks_tests_for_grouped_grades to perform Wilcoxon rank sum tests and estimate if differences in the distributions between R0-2 and R3 grades are statistically significant.

Plot bar charts with fractal dimensions as DR biomarkers

• Run script_check_fractal_dimension_performance to evaluate raw fractal dimensions as potential biomarkers for DR. This script reproduce Figure 6 from our paper.

Plot bar charts with AUC values obtained with fractal measurements

• Run script_check_fractal_measurement_performance to train L1/L2 regularized logistic regression classifiers using fractal measurements for a given DR classification problem. If you setup the configuration file accordingly, it reproduces Figure 7.

ROC curve for proliferative DR screening

• Run script_train_kernelized_logistic_regression_classifier to train L1/L2 regularized logistic regression classifiers using the fractal features that you want. At the end this will plot a ROC curve showing performance.

Contributing

If you want to contribute to this project, please contact me by e-mail to jiorlando -at- conicet.gov.ar

Authors

José Ignacio Orlando - All coding - Webpage - Github - Twitter

See also the list of contributors who participated in this project.

License

This project is licensed under the MIT License - see the LICENSE.md file for details

Acknowledgments

• This work is partially funded by ANPCyT PICT 2014-1730.
• J.I.O. and H.L.M. are funded by doctoral scholarships granted by CONICET.
• We would also like to thank NK and CFK. They know why :-)