Most common chronic disease is Diabetes that results in different ailments, and Diabetic retinopathy (DR) is a most critical issue and besides, the most common reasons behind loss of vision among diabetic patients. For the treatment of the patient affected and prevent vision loss, Automated Detection of diabetic retinopathy is used at an early stage from the massive scale of retinal images that helps the ophthalmologist. The present technique (deep Multi Instance Learning (MIL)), uses DR detection to extract features from data and classifies it and achieves enhancement in the detection DR images. It deepens the lesions. However, in the mass DR screening, the retinal images having lesser image quality will be surely required. In the recent techniques, the low illumination images may not be taken into consideration and the dark dots resulting due to the camera dust may be misclassified to be DR lesions. In order to mitigate these problems; the author introduced a novel quality evaluation metric for assessing the image quality. A structural similarity index metric is required for executing this task. New fuzzy histogram equalization is used for equalization and enhancement of contrast, in this newly introduced work. Depending on the brightness perceived of global diabetic retinopathy images, the quantity of saturated retina pixels is intelligently fixed. Then using the preprocessed image, the extraction of the image patches was often done. Then, it is given to the improved deep patch-level classifier, for the estimation of their Diabetic Retinopathy (DR) probabilities and the improved particle swarm optimization (IPSO) algorithm helps in optimizing the weights values of this classifier. In order to update the CNN's parameters for increasing the performance of detection, the improved particle swarm optimization (IPSO) and back propagation is used.
Volume 11 | 10-Special Issue
Pages: 494-505
DOI: 10.5373/JARDCS/V11SP10/20192836