Retinal image processing and classification using convolutional neural networks
- MSC MI 2016-2018
In the field of medical image processing, deep learning approaches is emerging out as powerful tool for analysis and disease detection. The project aims to develop a system to distinguish retinal diseases from fundus images. Precise and programmed analysis of retinal images has been considered as an effective tool for the determination of retinal diseases such as diabetic retinopathy, hypertension, arteriosclerosis, etc. In this study, we extracted different retinal features such as blood vessels, optic disc, lesions, etc., and applied Convolutional Neural network based models to detect various disease from fundus images involved in the STructured Analysis of REtina. Augmentation techniques like translations and rotations were done for increasing the number of images. The blood vessel extraction was done with the help of morphological operations like dilation and erosion, and enhancement operations like CLAHE, AHE. The bright lesions (exudates) inside the retina can be detected by the filtering operations and contrast enhancement after the removal of the optic disc. The optic disc can be localised by the methods such as opening, closing, Canny’s edge detection, and finally thresholding the image after filling the holes. At the end of the stage of classification, we will get the disease classes with the probabilities. We used the Convolutional Neural Networks for this grouping purpose. The convolutional model for 10 classes with the RGB images as input, got 42% as the accuracy. The model with the green channel images are expected to have the highest accuracy. Due to the small size of datasets, the deep learning techniques in this study were ineffective to be applied for the detection of the disease.
CONVOLUTIONAL NEURAL NETWORK RETINAL IMAGE PROCESSING DEEP LEARNING