Transfer Learning with Fine-Tuned MobileNetV2 for Diabetic Retinopathy

Blindness is a vision impairment that cannot be corrected fully with medication or surgery or through glasses. One of the reason for blindness is Diabetic Retinopathy. It is a medical condition that damages eye retinal tissues. In todays era emphasis is on finding automatic computational mechanism that can check the severity of diabetic retinopathy so that the blindness can be detected before it happened. Instead of designing a deep neural network from scratch this paper proposes an approach based on transfer learning. MobileNetv2, a predefined model is used for extracting a meaningful features from the given set of retina images. Model is customized by adding the globalaveragepooling layer and softmax classifier layer on the top of pretrained base model for classifying images in one of the five different classes of diabetic retinopathy. Initially we have only trained the stacked layers, thereby refraining the weights of pretrained model from updation. To further improve the performance of the model the weights of top few layers of the base network are fine-tuned. The network now adapts to itself with these specialized features. Kaggle diabetic retinopathy dataset is used for evaluating the performance of the proposed approach. 2929 retinal fundus training images, 733 validation images are used to build the model and tested with 1928 testing images. Experimental result shows that after fine tuning the network training accuracy increased from 70% to 91% and validation accuracy increased from 50% to 81%. Training loss and validation loss is observed to be approximately same, that indicates model is perfectly fit. This accuracy of fine-tuned network reveals a noticeable improvement.