Skip to content
RadIO is a library for data science research of computed tomography imaging
Branch: master
Clone or download
roman-kh Merge pull request #29 from henrique/patch-2
 minor: rename cv_split to split
Latest commit 6d53e25 Dec 18, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information.
examples minor: rename cv_split to split Dec 17, 2018
radio fix several more pylint errors Dec 3, 2018
.gitattributes Renamed .gitattributes Mar 13, 2017
.gitmodules add batch-flow sub Nov 14, 2018
LICENSE Remove unneeded files Jan 21, 2018 Update Dec 19, 2017
requirements.txt update augmentation-nb; add commits from kmader; update dataset-version Apr 11, 2018


RadIO is a framework for data science research of computed tomography (CT) imaging.

Main features:

  • Asynchronously load DICOM and MetaImage (mhd/raw) files
  • Dump files to blosc to compress datasets and thus accelerate loading
  • Transform and augment CT-scans in parallel for faster preprocessing
  • Create concise chainable workflows with actions or use tailored pipelines for preprocessing or model training
  • Train with ease a zoo of state-of-the-art neural networks for classification or semantic segmentation
  • Sample crops of any size from CT-scans for comprehensive training
  • Customize distribution of crop locations for improved training
  • Predict on the whole scan

The documentation contains a comprehensive review of RadIO's capabilities. While tutorials provide ready-to-use code blocks and a practical demonstration of the most important RadIO features.


There are four tutorials available:

  • In the first one you can learn how to set up a dataset of CT-scans and define a basic preprocessing workflow.
  • The second tutorial contains in-depth discussion of preprocessing and augmentation actions.
  • The third tutorial explains how to generate batches to train a neural network.
  • The fourth tutorial will help you configure and train a neural network to detect cancer.

Preprocess scans with chained actions

Preprocessing-module contains a set of actions to efficiently prepare a dataset of CT-scans for neural networks training.

Say, you have a bunch of DICOM scans with varying shapes. First, you create an index and define a dataset:

from radio import CTImagesBatch
from radio.batchflow import FilesIndex, Dataset

dicom_ix = FilesIndex(path='path/to/dicom/*', dirs=True)            # set up the index
dicom_dataset = Dataset(index=dicom_ix, batch_class=CTImagesBatch) # init the dataset of dicom files

You may want to resize the scans to equal shape [128, 256, 256], normalize voxel densities to range [0, 255] and dump transformed scans. This preprocessing can be easily performed with the following pipeline:

pipeline = (
    .resize(shape=(128, 256, 256))

See the documentation for the description of preprocessing actions implemented in the module.

Preprocess scans using a pre-defined workflow

Pipelines module contains ready-to-use workflows for most frequent tasks. For instance, if you want to preprocess a dataset of scans named dicom_dataset and prepare data for training a neural network, you can simply execute the following pipeline creator (without spending much time on thinking what actions to choose for a workflow):

from radio.pipelines import get_crops

nodata_pipeline = get_crops(fmt='raw', shape=(128, 256, 256),
                            nodules=nodules, batch_size=20,
                            share=0.6, nodule_shape=(32, 64, 64))

dicom_pipeline = dicom_dataset >> nodata_pipeline

for batch in dicom_pipeline.gen_batch(batch_size=12, shuffle=True):
    # ...
    # train a model here

See pipelines section for more information about ready-to-use workflows.

Adding a neural network to a workflow

RadIO contains proven architectures for classification, segmentation and detection, including neural networks designed specifically for cancer detection (e.g. DenseNoduleNet inspired by the state-of-the-art DenseNet, but well suited for 3D CT scans):

from radio.preprocessing import CTImagesMaskedBatch as CTIMB
from radio.models import DenseNoduleNet
from radio.batchflow import F

training_pipeline = (
      .sample_nodules(nodule_size=(32, 64, 64), batch_size=20)
      .init_model('static', DenseNoduleNet, 'net')
      .train_model('net', feed_dict={
          'images': F(CTIMB.unpack, component='images'),
          'labels': F(CTIMB.unpack, component='classification_targets')
), shuffle=True)

The models documentation contains more information about implemented architectures and their application to cancer detection.


RadIO module is in the beta stage. Your suggestions and improvements are very welcome.

RadIO supports python 3.5 or higher.

Installation as a python package

With pipenv:

pipenv install git+

With pip:

pip3 install git+

After that just import radio:

import radio

Installation as a project repository:

When cloning repo from GitHub use flag --recursive to make sure that BatchFlow submodule is also cloned.

git clone --recursive

Citing RadIO

Please cite RadIO in your publications if it helps your research.


Khudorozhkov R., Emelyanov K., Koryagin A., Kozhevin A. RadIO library for data science research of CT images. 2017.
  author = {Khudorozhkov R., Emelyanov K., Koryagin A., Kozhevin A.},
  title  = {RadIO library for data science research of CT images},
  year   = 2017,
  doi    = {10.5281/zenodo.1156363},
  url    = {}
You can’t perform that action at this time.