• Kalyan K Mohanty

Prediction of diabetic from retinopathy using machine learning and signal processing approach

Diabetic retinopathy disease is constantly on the rise across the globe which causes blurry vision, partial and total blindness to diabetic affected people. Advancement of Biomedical imaging with signal processing and machine learning algorithms make ease of the prediction of Diabetic retinopathy to a greater extent. In this study, the retinal image is taken from a fundus camera of both healthy and diabetic retina. Image pre-processing techniques, morphological operations are used to detect the statistical features and the histogram-based feature is extracted by using Discrete Wavelet Transform (DWT) which is the novel contribution of the proposed algorithm. These features are classified by any machine learning approach (K-Nearest Neighbors, Support Vector Machine and Artificial Neural Network) to predict DR accurately and efficiently following a cross-validation approach.

Diabetic Retinopathy can be divided into two types i.e. Non-Proliferative Diabetic retinopathy (NPDR) and Proliferative Diabetic Retinopathy (PDR). NPDR is an early stage and can be detected by the presence of microaneurysms and if it’s not properly taken care then it leads to its advance stage PDR. Lack of oxygen in the retina occurs when the disease progresses by time which forms new blood vessels and makes clouding in the patient’s sight. With the help of a fundus camera, microaneurysms and exudates can be detected. They appear as the red dot and white or yellow spots respectively. Exudates are classified into two types as hard exudates which has well-defined boundaries and soft exudates having blurry boundaries.

(In this research, 196 JPG images of diabetic and non-diabetic patients taken by the fundus camera of resolution 760 x 570 pixels.)

Image Pre-processing:

The raw image should go through the preprocessing stage because in this stage image gets rectified; problems such as blurring, resizing, etc. are done. RGB images are taken as input but to get the effective results it needs to be converted to HSV (Hue saturation Value) image by the help of color space conversion or green channel extraction can be also very helpful for intensity conversion from RGB image. Further image enhancement can be done by histogram equalization.

Blood vessel extraction:

In the retinal image, microaneurysms have a similar concentration of blood vessels. So it is essential to extract blood vessels from the DR image to obtain an accurate result.

Feature extraction:

When there are considerably huge data with many redundant features then researchers have developed some techniques to get rid of those redundant data and make them useful as transforming as a useful set of data is called feature extraction. So the given task can be performed with a minimum no of data.

Statistical data extraction:

Image after gone through pre-processing and blood vessel extraction then the data is ready to extract from the set of the image by calculating their entropy, energy, standard deviation, mean, variance, contrast, etc. which is called statistical data extraction. Five statistical features are extracted for better prediction accuracy.

Discrete Wavelet Transform (DWT):

DWT data extraction Wavelet transform is one of the mathematical tools which gives allows the researcher to analyze the various data of images with multiple resolutions. Smooth regions of image interrupted by edges or abrupt changes. A most important part of the data is present in these abrupt changes. The FT (Fourier Transform) is a tool that can be used for data analysis. However, the abrupt changes can’t be represented efficiently in Fourier transform. The reason for that FT represents the data as the sum of sine waves that are not localized in both time and frequency. This sine wave oscillates forever. Therefore to accurately analyze signal and image we need WT which is localized in both time and frequency. In discrete wavelet analysis, the information is not repeated which is stored in the wavelet coefficient, it allows the complete regeneration of the signal without having redundancy. This property has inspired the development of wavelet-based signal compression. For image denoising, Wavelet thresholding techniques are generally used which gives the ability of wavelet analysis to distinguish noise from the image signal. However, choosing a suitable mother wavelet gives better efficiency.


A. KNN Classifier

KNN algorithm is used for both regression and classification. This algorithm is simplest among all other machine learning algorithms. The distance between a test and a training sample is generally based on Euclidean distance

B. SVM Classifier

SVM algorithm is used for supervised machine learning that is both for data classification and regression analysis. It classifies efficiently both linear and nonlinearly separable data by creating a hyper-plane for data in the original space.

C. ANN Classifier

There are three basic blocs in ANN by which classification is carried out. All key features of a signal are extracted in the pre-processing stage. Weight and bias optimization is done in the learning or training phase. Finally, in test phase decisions are taken by the neural network


The experimental result shows that the obtained accuracy and other measures are better for DWT with histogram features, while the statistical features have a very lower value. While SVM and ANN performed better in case of statistical features but the ANN performance is better as compared to the KNN and SVM for DWT features. However, as a whole, the DWT based histogram-based features are suggested for the diabetic retinopathy classification of fundus images.

Link to paper:

Link to github: