Skip to content

Machine learning models for multi-organ, multi-disease prediction in chest CT volumes. From paper Draelos et al. "Machine-Learning-Based Multiple Abnormality Prediction with Large-Scale Chest Computed Tomography Volumes."

License

Notifications You must be signed in to change notification settings

rachellea/ct-net-models

Repository files navigation

ct-net-models

Logo

Description

This repository contains Python code to train and evaluate convolutional neural network models (CNNs) on the task of multiple abnormality prediction from whole chest CT volumes. Our model CT-Net83 achieves state of the art performance on this task and is described in detail in our Medical Image Analysis paper. The paper is also available on arXiv. The models are implemented in PyTorch.

On the RAD-ChestCT data set of 36,316 volumes from 19,993 patients, CT-Net83 achieves a test set AUROC >0.90 for 18 abnormalities, with an average AUROC of 0.773 across 83 abnormalities.

If you find this work useful in your research, please consider citing us:

Draelos R.L., et al. "Machine-Learning-Based Multiple Abnormality Prediction with Large-Scale Chest Computed Tomography Volumes." Medical Image Analysis (2020).

Requirements

The requirements are specified in ctnet_environment.yml and include PyTorch, numpy, pandas, sklearn, scipy, and matplotlib.

To create the conda environment run:

conda env create -f ctnet_environment.yml

The code can also be run using the Singularity container defined in this repository.

Usage

To run a demo of the CT-Net83, CT-Net9, BodyConv, 3DConv, and ablated CT-Net models on fake data, run this command:

python main.py

The RAD-ChestCT data set is publicly available on Zenodo.

Because the real dataset is large, currently this repository includes fake data files to enable demonstrating the code and the required data formats. The fake data is located in load_dataset/fakedata. The fake CTs were generated as follows:

fakect = np.random.randint(low=-1000,high=1000,size=(10,10,10))
np.savez_compressed('FAKE000.npz',ct=fakect)

Note that 10 x 10 x 10 is too small for a real CT scan; the CT scans in the RAD-ChestCT data set are on the order of 450 x 450 x 450 pixels.

Organization

  • main.py contains the experiment configurations needed to replicate the results in the paper. The command python main.py will run a demo of the CT-Net83, CT-Net9, BodyConv, 3DConv, and ablated CT-Net models on fake data.
  • run_experiment.py contains code for training and evaluting the models.
  • evaluate.py contains code for calculating, organizing, and plotting performance metrics including AUROC and average precision.
  • unit_tests.py contains some unit tests. These tests can be run via python unit_tests.py
  • load_dataset/custom_datasets.py contains the PyTorch Dataset class for the CT data.
  • load_dataset/utils.py contains the code for preparing individual CT volumes, including padding, cropping, normalizing pixel values, and performing data augmentation through random flips and rotations.
  • load_dataset/fakedata contains the fake data necessary to run the demo.
  • models/custom_models_ctnet.py contains the CT-Net model definition.
  • models/custom_models_alternative.py contains two alternative architectures, BodyConv and 3DConv.
  • models/custom_models_ablation.py contains the ablated variants of CT-Net.

Comment on Data Parallelism

Currently the experiments in main.py are set up to replicate the paper results. Several of the experiments use data parallelism which assumes that at least 2 GPUs are available. If you wish to run the demo on one GPU, then change batch_size to 1 and set data_parallel to False.

Logo

The logo includes two Creative Commons icons from the Noun Project: lungs and gear.

About

Machine learning models for multi-organ, multi-disease prediction in chest CT volumes. From paper Draelos et al. "Machine-Learning-Based Multiple Abnormality Prediction with Large-Scale Chest Computed Tomography Volumes."

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages