Retinostic is able to diagnose extremely early stages(no symptoms) of Diabetic Retinopathy from retinal images. This is a vision threatening disease affecting about 347-million people with diabetes.
Retinostic is a Residual block based deep neural network that analyzes images of the retina to categorize them into various levels of diabetic retinopathy. This is especially useful in rural areas where advanced/sophisticated medical diagnosis devices are not available. This model has an accuracy of about 96.4% based on the performance with open source visual data from Kaggle.
ResNets are quite powerful and complex neural networks and can be built in a limited amount of time. The data was separated into training, validation, and testing data in my program.
My algorithm takes an image, applies convolutions--Conv2D--(extracts important features from the image e.g., hue variation, white spotting on the retinal surface etc), creates feature maps, does pooling/downsampling (reducing the size of the feature maps), does flattening(turns 2d pixels to 1d), and then feeds the feature maps to the dense neural networks.
ResNets aka Residual Neural Networks add a benefit to CNNs by solving the vanishing gradient problem using skip connection. The algorithm is quite extensive with multiple blocks and layers involved as you'll see in the iPython file attached below.
I am using Keras and Tensorflow in my deep learning model that ended up having just under 5 million trainable parameters!
In conclusion, this model has a lot of scope in the diagnosis of diabetic retinopathy.