Kaggle Diabetic Retinopathy Detection

This is a CNN detector for the Kaggle DR dataset. It takes retinal fundus photography as input, and predicts DR severity (0-4).

Origin and Network Archiecture

This project is derived from sveitser/kaggle_diabetic, which was team o_O's solution for the Kaggle Diabetic Retinopathy Detection Challenge.

The original implementation relies on Python2 and some obsoleted deep learning frameworks. I have modified it so that it can be run on Python3 + keras + tensorflow frameworks. Please note that this implementation is different from the original on several points:

The left eye and right eye are discoupled and used as separate training sample.
The original blending procedure has not been implemented.

Please Refer to the original solution report for more information on the network and training architecture.

Reqirements

This project is built on Python3 + Tensorflow1.x and Keras2.x. The exact dependencies are listed in requirements.txt.

pip3 install -r requirements.txt

Dataset Preparation

Download Kaggle DR dataset and extract train/test images to data/train and data/test respectively, and put the trainLabels.csv and retinopathy_solution.csv label file into the data directory as well.

convert.py will do center-cropping and resizing to the original Kaggle dataset. The o_O solution require training images of 3 different sizes, 512x512, 256x256 and 128x128.

python3 convert.py --crop_size 512 --convert_directory data/train_medium --extension jpg --directory data/train
python3 convert.py --crop_size 512 --convert_directory data/test_medium --extension jpg --directory data/test

python3 convert.py --crop_size 256 --convert_directory data/train_small --extension jpg --directory data/train_medium
python3 convert.py --crop_size 128 --convert_directory data/train_tiny --extension jpg --directory data/train_medium

Start Training

The convolutional network configuration is done via the files in the configs directory. You can change parameters such as learning rate, batch size, and can also modify the network.

The o_O solution first trains model of smaller input images, and use them to initalize models of larger input images. For example runing these lines to get a 512_4x4_32 model:

# Train network with 4x4 kernels.
python3 train_nn.py --cnf configs/c_128_4x4_32.py
python3 train_nn.py --cnf configs/c_256_4x4_32.py --weights_from weights/c_128_4x4_32/weights.h5
python3 train_nn.py --cnf configs/c_512_4x4_32.py --weights_from weights/c_256_4x4_32/weights.h5

After training, you can find a best-val_kappa model weight at weights/c_512_4x4_32/weights.h5.

A pretrained model can be found here: o_O-DR-detector-c_512_4x4_32.h5.

Evalation

Run eval_nn.py to get the performance of a model on the Kaggle test set.

python3 eval_nn.py --cnf configs/c_512_4x4_32.py --weights_from weights/c_512_4x4_32/weights.h5

Example output:

kappa:  0.77537376
accuracy: 0.7821039271315514
             precision    recall  f1-score   support

          0       0.90      0.91      0.90     39533
          1       0.17      0.27      0.21      3762
          2       0.72      0.52      0.60      7861
          3       0.34      0.44      0.39      1214
          4       0.83      0.35      0.49      1206

avg / total       0.81      0.78      0.79     53576

Testing

Run test_nn.py to get predicted labels on user-designated images. For example you've put the resized images into a path_to_test_images/ directory, run

python3 test_nn.py --cnf configs/c_512_4x4_32.py --weights_from weights/c_512_4x4_32/weights.h5 --input_dir path_to_test_images/