Feature Extraction From Retinal Fundus image for automated grading of Diabetic Retinopathy
This is a short and brief abstract to the work i am doing.
Before i begin i would like to thank Yash,Dinesh,Madan and Nistha of Biomedical Science Department at ANDC for helping me in understanding the biology behind the disease.
I would also like to thank Dr Sukanto Deb and Dr Subhash Kumar who tought me Fourier transform which i implemented during the feature extraction process.
INTRODUCTION:- Diabetic Retinopathy is an eye disease that happens to people having long-standing diabetes. Currently it is detected by practitoners by analysing retinal fundus image manually, it is a time consuming process thus delaying the treatment and also constraining the reach to economically lower class of population. Therfore there is prime need to device automated system to grade Diabetic Retinopatahy. This can be fulfiled by making machine learning models based on supervised learning of lebeled fundus images.
AIM:- To build an automated Classifier to classify retinal fundus images and grade them on the scale of 0 to 4 where 0 corresponds to no sign of diabetic retinopathy and 4 corresponds to severe case of diabetic retinopathy.
METHODS BEING APPLIED:-
We need to extract the relevant features from the image before we apply any machine learning technique to classify the image.
1)Image Processing:- Various image processing algorithm has been applied in a sequential manner to extract the relevant information from the data. Which in turn will form as the feature to the classifier at the latter stage.
a) Blood Vessel Extraction:- Blood vessels forms one of the main feature for the latter classification. They are 1st extraced from the image by using different image processing operations applied systematically and then they are quantified to extract some features. In our case we extracted(Area,Homogenity,Entropy etc) from the image of blood vessels + bright exudates that we extracted.
b)Optic Disc Extraction:- Optic disc extraction is also one of the initial steps for diabetic retinopathy grading using fundus image. Area of the optic disc was used as a feature for our classifier and also masking of Optic disc was necessary for locating and extracting hard exudates.
c) Fovea localisation:- Fovea localisation is a very important step for grading diabetic retinopathy as the distance of exudates from fovea is one of the most important feature for grading DR.
Feature Extraction from an image in itself is a very tough work, and automating each extraction process so that it can be be applied to 40,000 images which are taken in variable light condition, makes it a more challenging work and thus a very interesting work to learn from.
This is also a well studied disease and so the ammount of literature available isn't a problem thus a good thing for beginners to try their hands on.
A glimpse of the features extracted so far.
a) area:- It denotes the area covered by blood vesseles.
b) contrast,correlation,energy,homegenity are the various features extracted after the texture analysis of the image containing only blood vessels and exudates.
c) area_optic:- It represents the area of optic disc.
Further I'm currently working on to extract microaneurysm using fourier transform and also correlate the distance of exudates from fovea to make a new feature.
Simultaneously i am trying different machine learning models to find the best model for my extracted data.
Comentarios