Classification of Diabetic Retinopathy Using Few-Shot Transfer Learning from Imbalanced Data

Diabetic Retinopathy (DR) has no effective cure if it has reached the proliferative phase and blindness will most probably be cursed on patients ultimately. Good prognosis depends on early detection of DR and Glaucomatous. In this paper, three types of categorical data were fed into AlexNet and VGG16 models. Primarily mild disease was misclassified as the convolutional neural network struggles to distinguish delicate features of disease among normal and mild phase. To mitigate erroneous impact, various image preprocessing techniques such as otsu thresholding, global histogram equalization (GHE), contrast limited adaptive histogram equalization (CLAHE) etc. have been used for the recognition of subtle features lying in fundoscopic images. Pretrained architecture VGG16 model with no freezing of parameters achieved good fitting and accuracy over 3 ary classification of small balanced dataset. The AlexNet achieved better fitting over VGG16 for both 2 ary and 4 ary classification with an optimized learning rate with imbalanced dataset. The deeper VGG16 exploits deeper features from each class thus it faces biased effect from weightier class. Thus, VGG16 performs better over 3 ary dataset through deeper features exploitation. AlexNet model has achieved better results through batch-normalization and batch-wise training methods over a slightly imbalanced dataset.