The goal of this project is to develop high-accuracy deep learning models for identifying 14 thoracic diseases from chest X-ray scans, as well as to localize the regions of the scans indicating disease.

Here is a paper summarizing our methods and findings.

Notes:

• Unless stated otherwise, run all scripts from the top-level folder of this project.
• We recommend running this project from within an Amazon Web Services (AWS) Deep Learning Base AMI (Ubuntu), using a GPU-accelerated instance type such as p2.xlarge.
• To learn about the user arguments for each script, run [script name].py -h.

Ensure that pip and Anaconda are installed. Then run:

conda install pytorch torchvision -c pytorch
pip install -r requirements.txt

The NIH hosts the dataset at this Box folder. Download all the tar.gz files in images, and extract them to the folder /data/images, so that we have:

data/
    images/
        00000001_000.png
        00000001_001.png
        00000001_002.png
        ...
        (112,120 images in total)

Run the script build_dataset.py, which resizes the images to 224x224 and stores them in data/224x224_images. (The new image size, source directory, and output directory can be adjusted via user arguments.)

python build_dataset.py

Run the script train.py, which trains a DenseNet169 model on the dataset using the hyperparameters specified in experiments/base_model/params.json, then stores the parameters of the model in a file.

python train.py

Run the script analyze_feature_vectors.py, which runs the dataset through the trained DenseNet model and writes the feature vectors (defined as the output of the second-to-last layer of the DenseNet) to a file. It also prints information on the average distances between vectors in the dataset.

python analyze_feature_vectors.py

Run the script classify_by_cluster.py, which uses the feature vectors extracted from the DenseNet as inputs to various traditional machine learning models. By default, this script trains and evaluates (on the validation set) a random forest model and a k-nearest neighbors model.

python classify_by_cluster.py

Run the script evaluate.py, which loads the DenseNet169 model trained previously, and evaluates it on the test set.

python evaluate.py

Run the script classify_by_cluster.py with an appropriate argument to use the test set instead of the validation set:

python classify_by_cluster.py --dataset_type test

Run the script evaluate_ensemble.py, which loads the DenseNet169 model trained previously, trains any number of non-neural machine learning models on the feature vector set, and evaluates the ensemble of these models on the test set. By default, only non-neural model, namely a random forest, is used in this ensemble.

python evaluate_ensemble.py

Under the experiments directory you will find a file called params.json, which contains baseline hyperparameter settings. It looks like:

{
    "learning_rate": 1e-4,
    "learning_rate_decay_factor": 0.2,
    "learning_rate_decay_patience": 1,
    ...
}

Suppose we wish to search over different learning rates. Create a new directory under experiments, say experiments/learning_rate, and create a similar params.json file under that directory. Then in search_hyperparams.py, uncomment the code block that creates a training job for each learning rate; this block looks like:

# Perform search over learning rate
learning_rates = [1e-5, 1e-4, 1e-3, 1e-2]

for learning_rate in learning_rates:
    # Modify the relevant parameter in params
    params.learning_rate = learning_rate

    # Launch job (name has to be unique)
    job_name = "learning_rate_{}".format(learning_rate)
    launch_training_job(args.parent_dir, args.data_dir, small_flag, job_name, params)

Now run search_hyperparams.py. The results for each choice of learning rate will be stored in a directory under experiments/learning_rate.

To display the results of this hyperparameter search in a table, run

python synthesize_results.py --parent_dir experiments/learning_rate

Currently, there is no automated script to perform hyperparameter search for non-neural models. Instead, manually adjust the code in the main() function of classify_by_cluster.py and then run that script.