APTOS 2019 Blindness Detection: 37th Place(Top 2%) Solution

Aravind Eye Hospital in India provided a large set of retina images taken using fundus photography under a variety of imaging conditions. A clinician has rated each image for the severity of diabetic retinopathy on a scale of 0 to 4. The task of this competition was to train a network that can detect the level of severity.

Our approach

I treated the task as a regression problem. I experimented with several Resnet, SE-Resnet, Resnet, ResNext and SE-Resnext models until I found out that Efficientnet model performs better than most of the pretrained models in this competition. Our final silver winning solution was an ensemble of 5 different Efficientnet models. Here is an overview:

• Preprocessing: Only Circle Cropping from this kernel.

• Augmentations: Horizontal and Vertical Flip, 360 rotation, zoom 20%-25%, lighting 50%. Used Fast.ai transformations.

• Training: Mostly followed this method. Did 5+10 training using 2015 data as train and 2019 as validation set. Then I split the 2019 data into 80:20 ratio using sklearn's Stratified Split function and finetuned the previously trained model over it for 15 epochs. I took the best model based on validation loss which didn't seem to be overfitted and used it for submission. I used a fixed seed so that I can compare different model performances over the same val set. Didn't use CV or TTA. I wrote a pytorch and a fastai version of my codes for the competition.

I mostly used my fastai codes throughout the competition. It requires a lot of work to cleanup my codes and add a detailed explanation of my workflow. I guess I will keep updating this repo.

Approaches That Didn't Work For Me

Classification and Ordinal Regression

Scores (TL;DR)

Here is the summary of my final models and their ensemble scores:

Model	Image Size	Val Kappa	Public Kappa	Private Kappa
Efficientnet B2	256	0.922	0.807	0.918
Efficientnet B1	256	0.926	0.804	0.921
Efficientnet B0	256	0.919	0.816	0.914
Efficientnet B3	256	0.921	0.812	0.917
Efficientnet B5	300	0.920	0.802	0.916
Ensemble	----	0.932	0.826	0.926

Resources

• Competition Link
• Pytorch Sample Code
• Weighted Kappa and Metric Optimizer Link
• Data augmentation library Albumentation example
• Preprocessing Technique (borrowed from last competition winner)
• Discussion on our strategy.
• All gold medal solutions and discussions

Requirements

• torch
• torchvision
• fastai
• sklearn
• numpy
• tqdm

PyTorch

• Steps
  • Create a directory data/new_data and keep the 2019 APTOS competition data there.
  • Run cyclic_lr.py and find a suitable learning rate.
  • Run cd pytorch_version and then train.py. It will run a 5 fold cross-validation with 10 epochs for each fold.
  • You can experiment with albumentations augmentations by using DRDatasetAlbumentation as your Dataset class.
  • The code monitors both val_loss and kappa scores and saves model based on them. In my experience, kappa score is unstable and often doesn't seem to be correlated with val_loss. The safe option here is to choose and save your model based on the val_loss only.

Fastai

The notebook fastai_version/Blindness_Detection.ipynb was written for training a model over both 2015 and 2019 datasets. This notebook contains a complete set of instruction and codes from downloading the datasets to train a model in 2 phases.

• Steps

Will be added soon