Diabetic retinopathy Detection
2nd Prize Kaggle DiaBetic retinopathY Competition
Summary
We use VGG style convolutional neural networks trained with "Lasagne" and "nolearn" with large color images, various types of data augmentation, dynamic resampling for class imbalance and a "per patient'" feature blending strategy which takes advantage of having pairs of sample images for each patient at our disposal. The final solution is a simple average of blends with features from two deep convolutional networks and three sets of weights for each network.