A Feture Extraction Approach for the Computational Diabetic Retinopathy Detection in Retinal Images

Remark, that GPU and CPU related proceedings were accomplished on a linux based machine.

Acknowledgement

I like to acknowledge Dr. Manfred Liebmann from the University of Graz for guiding me through the process of creating this software.

Software Requirements

•	Matlab
•	CUDA Toolkit (Version >= 7.0)

Hardware Requirements

•	NVIDIA GPU with CUDA
    	Minimum Architecture: 2.0 (not tested / makefile has to be adapted)
    	Suggested Architecture: >= 3.5
•	Modern CPU

Data Resource

•	Jorge Cuadros and George Bresnick Eyepacs: An adaptable telemedicine system for diabetic retinopathy screening Journal of diabetes science and technology (Online), 3(3):509–516, 05 200
•	Further information: https://www.kaggle.com/c/diabetic-retinopathy-detection

Data Format and Structure

•	88702 JPEG Images
•	Feature Extraction on GPU
	⁃	Folder with Subfolders
	⁃	Each Folder represents a particular Stage of DR (e.g. ../RetinalImages/0/ ../RetinalImages/1/ ../RetinalImages/2/ ../RetinalImages/3/ ../RetinalImages/4/)
	⁃	Retinal Grey Scale Images (Suggested: Green Component) are within the corresponding Folder 
	⁃	.txt File with labels holding following Format:
		10_left,10_right,11_left,11_right,…,30000_left
	⁃	A prior matching evaluates whether a particular label is contained within the folder, hence number of labels do not have to match number of images in the referrenced folder
•	Feature Extraction on CPU
	⁃	Folder with Subfolders
	⁃	Each Folder represents a particular Stage of DR (e.g. ../RetinalImages/0/ ../RetinalImages/1/ ../RetinalImages/2/ ../RetinalImages/3/ ../RetinalImages/4/)
	⁃	Retinal Grey Scale Images (Suggested: Green Component) are within the corresponding Folder 

SVM Approach / Feature Extraction

•	Matlab / CPU
	⁃	Primary used for comfortable Visualization of Results and Establishment of Image Processing Pipeline
	⁃	Open File ../CPU/featureExtraction.m in Matlab
	⁃	Execute
			function [ MAT, rowValue ] = featureExtraction(imagePath, class, destinationPath, plot)

	⁃	Corresponding Explanation of Variables is stated within the m-File

•	CUDA / GPU
	⁃	Compile Software in Folder ./GPU/ with enclosed makefile (nvcc compiler required)
	⁃	Execute: ./featureExtraction
	⁃	Exemplary Run:
		===================================================================
		Image Processing Library
		Version: 1.0
	
		On this system are 4 CUDA devices available.
		Device: Tesla K20m  Streaming MP: 13  Compute Version: 3.5
	
		Directory to image resource: /home/kapsecker/imagesGrey/3/
		Directory to image destination: /home/kapsecker/images/3/
		Directory to labels: /home/kapsecker/res/labels.txt
		-------------------------------------------------------------------
		0/2086 : /home/kapsecker/imagesGrey/3/99_left.jpeg
		1/2086 : /home/kapsecker/imagesGrey/3/99_right.jpeg
		2/2086 : /home/kapsecker/imagesGrey/3/163_left.jpeg
		…
		2085/2086 : /home/kapsecker/imagesGrey/3/44333_left.jpeg
		2086/2086 : /home/kapsecker/imagesGrey/3/44333_right.jpeg
		-------------------------------------------------------------------

		Elapsed time: 135.79 seconds.
		Finished successfully!
		===================================================================

SVM Approach / Classification:

•	Resulting Feature Matrix is saved in following Form

		Image,Class,Bloodvessels,Haemorrhages,Exudates,Contrast,
		/home/kapsecker/imagesGrey/2/30_right.jpeg,2,0.000616252,0,0,0.255243,
		/home/kapsecker/imagesGrey/2/40_left.jpeg,2,4.10808e-05,0,0,0.273186,
		/home/kapsecker/imagesGrey/2/51_left.jpeg,2,4.98195e-06,0,0,0.116361,
		/home/kapsecker/imagesGrey/2/54_left.jpeg,2,0.00131854,0,0,0.175058,
		/home/kapsecker/imagesGrey/2/78_left.jpeg,2,0.000438692,0,0,0.114433,
		/home/kapsecker/imagesGrey/2/78_right.jpeg,2,0.000346607,0,0,0.109931,
		/home/kapsecker/imagesGrey/2/79_left.jpeg,2,4.25127e-06,0,0,0.191255,
		/home/kapsecker/imagesGrey/2/79_right.jpeg,2,8.10398e-06,0,0,0.270105,
		/home/kapsecker/imagesGrey/2/82_left.jpeg,2,0.00180359,0,0,0.650781,
		/home/kapsecker/imagesGrey/2/129_left.jpeg,2,8.80809e-05,0,0,0.135238,
		/home/kapsecker/imagesGrey/2/129_right.jpeg,2,0.000138166,0,0,0.148248,
		…

•	Import resulting Feature Matrix to Matlab’s Classification Learner
•	Train with desired Classification Learner

Results:

•	Feature Matrices, obtained from Feature Extraction Process, are saved in ../Features/
	Remark, that they are distinguished by Diabetic Retinopathy Classes

Edit / Redistribution:

•	MIT License
•	See License File