Classification of Diabetic Retinopathy using pre-trained Deep Learning Model- DenseNet 121
Seema Gulati, Kalpna Guleria, Nitin Goyal
Abstract
An upsurge in cases of diabetes patients is witnessed worldwide with 108 million sufferers in the year 1980 to a whopping 422 million in the year 2014. The growth rate is lesser in countries with high earnings in comparison to the nations with small and intermediate earnings. Diabetes causes many other ailments in the body like kidney malfunction or failure, permanent vision loss, cardiac arrest, strokes, and amputation of the lower limb. Generally, the populace having diabetes mellitus for a prolonged period like that of more than 15 years have a higher probability of developing diabetic retinopathy (DR) that damages the eyes permanently if not treated in the earlier stages. The new age automatic detection system employing deep learning (DL) algorithms can prove to be fruitful in detecting diabetic retinopathy in the initial stages and therefore treatment can be provided well in time. In this paper, one of the most coveted DL techniques DenseNet is used to classify DR into five different stages. The hyperparameters are tuned and the maxima are found for these parameters, batch size, learning rate and epochs. These hyperparameters are usually application-specific and vary in accordance with the type and size of data.