English - 中文

Code for JAMA 2016; 316(22) Replication Study

Published article link: doi:10.1371/journal.pone.0217541.

Published trained neural network models: doi:10.6084/m9.figshare.8312183.

Abstract

We have attempted to replicate the methods and reproduce the results in Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs, published in JAMA 2016; 316(22) (link), using publicly available data sets. We re-implemented the main method in the original study since the source code is not available.

The original study used non-public fundus images from EyePACS and three hospitals in India for training. We used a different EyePACS data set from Kaggle. The original study used the benchmark data set Messidor-2 to evaluate the algorithm’s performance. We used another distribution of the Messidor-2 data set, since the original data set is no longer available. In the original study, ophthalmologists re-graded all images for diabetic retinopathy, macular edema, and image gradability. We have one diabetic retinopathy grade per image for our data sets, and we assessed image gradability ourselves.

We were not able to reproduce the original study’s results with publicly available data. Our algorithm’s area under the receiver operating characteristic curve (AUC) of 0.951 (95% CI, 0.947-0.956) on the Kaggle EyePACS test set and 0.853 (95% CI, 0.835-0.871) on Messidor-2 did not come close to the reported AUC of 0.99 on both test sets in the original study. This may be caused by the use of a single grade per image, or different data. This study shows the challenges of reproducing deep learning method results, and the need for more replication and reproduction studies to validate deep learning methods, especially for medical image analysis.

Requirements

Python requirements:

• Python >= 3.6
• Tensorflow >= 1.4
• OpenCV >= 1.3
• Pillow
• h5py
• xlrd
• matplotlib >= 2.1

Other requirements:

• p7zip-full
• unzip

Preprocessing before training

1. Run $ git submodule update --init to load the create_tfrecords repository. This tool will convert the data sets into TFRecord files.

2. Download the compressed Kaggle EyePACS data set and place all files (i.e. train and test set and labels) in the data/eyepacs folder. We recommend you to use the Kaggle API.

3. Run $ ./eyepacs.sh to decompress and preprocess the Kaggle EyePACS data set, and redistribute this set into a training and test set. Run with the --only_gradable flag if you want to train and evaluate with gradable images only. NB: This is a large data set, so this may take hours to finish.

For Messidor-2:

4. Run $ ./messidor2.sh to download, unpack, and preprocess the Messidor-2 data set. This data set is downloaded from the Datasets and Algorithms' section on Michael D. Abramoff's page here.

For Messidor-Original:

5. Download the Messidor-Original data set and place all files in the data/messidor folder.

6. Run $ ./messidor.sh to preprocess the Messidor-Original data set. Run with the --only_gradable flag if you want to evaluate with gradable images only.

Training

To start training with default settings, run $ python train.py. Optionally specify the path to where models checkpoints should be saved to with the -sm parameter.

Run $ python train.py -h to see additional optional parameters for training with your own data set, or where to save summaries or operating threshold metrics.

Evaluation

To evaluate or test the trained neural network on the Kaggle EyePACS test set, run $ python evaluate.py -e. To evaluate on Messidor-2, use the -m2 flag. To evaluate on Messidor-Original, run it with the -m flag.

To create an ensemble of networks and evaluate the linear average of predictions, use the -lm parameter. To specify multiple models to evaluate as an ensemble, the model paths should be comma-separated or satisfy a regular expression. For example: -lm=./tmp/model-1,./tmp/model-2,./tmp/model-3 or -lm=./tmp/model-?.

NB: The trained models used in this study are publicly accessible and can be downloaded from here.

The evaluation script outputs a confusion matrix, and specificity and sensitivity by using an operating threshold. The default operating threshold is 0.5, and can be changed with the -op parameter.

Run $ python evaluate.py -h for additional parameter options.

Evaluate on a custom data set

To evaluate the trained neural network on a different data set, follow these steps:

1. Create a Python script or start a Python session, and preprocess and resize all images to 299x299 pixels with resize_and_center_fundus from lib/preprocess.py.

2. Create a directory with two subdirectories image_dir/0 and image_dir/1. Move the preprocessed images diagnosed with referable diabetic retinopathy to image_dir/1 and the images without rDR to image_dir/0.

3. Create TFRecords: $ python ./create_tfrecords/create_tfrecord.py --dataset_dir=image_dir (run with -h for optional parameters).

4. To evaluate, run $ python evaluate.py -o=image_dir. Run with -h for optional parameters.

Benchmarks

We forked the repository for TensorFlow benchmarks (tensorflow/benchmarks) to run the benchmarks with the retinal fundus images and labels used in this study (link to fork). We further provide a Docker image maikovich/tf_cnn_benchmarks to run the benchmarks easily.

To run the benchmarks for training with this study's data on GPUs, run the following command with Docker. Substitute path/to/train_dir with the path to the directory that contains the TFRecord shards with the training set. The --num_gpu, --batch_size flags can be modified according to your environment's capabilities. For other flags, see the forked repository.

$ nvidia-docker run --mount type=bind,source=/path/to/train_dir,target=/data maikovich/tf_cnn_benchmarks:latest --data_dir=/data --model=inception3 --data_name=retina --num_gpus=2 --batch_size=64

As a reference, our training speed was 91.1 (± 0.4) images per second with one NVIDIA GTX 1080. For our training set of 57 146 images, that means approximately a minute per epoch of training. Add half a minute validation after each epoch due to initialization and loading into memory. Total training took typically 80 to 100 epochs, so it took around 2 to 2.5 hours to train one model.

To run the benchmarks to measure performance of evaluating test data sets, run the above command with the --eval flag and substitute the mount source directory with the path to the directory containing the TFRecord shards with test data.

In this repository, we also provide a example jinja file (see benchmarks.yaml.jinja.example) to generate a Kubernetes yaml description. This can be used to run the benchmarks in a distributed fashion on multiple nodes on a Kubernetes cluster. To see more information about how to compile the jinja file into yaml, see this README.

An overview of the original benchmark results with e.g. ImageNet data can be found here.