DABC-Net

DABC-Net toolkit is designed for fast and robust pneumonia segmentation and prediction of COVID-19 progression on chest CT scans. The core of the toolkit, DABC-Net, is a novel deep learning (DL) network that combines a 2D U-net for intra-slice spatial information processing, and a recurrent LSTM network to leverage inter-slice context. Compared to other popular volumetric segmentation networks such as 3D U-net, DABC-Net is much faster and more robust to CT scans with various slice thickness.

Based on DABC-Net segmentation, we can predict the disease progression, i.e. whether a specific patient will develop into a severe stage or not using his/her first two CT scans. This repository provides an implementation of DABC-Net (including graphical user interface), which can be potentially used to support early triage of severe patients

The main features:

Table of Contents

Installation

If you run toolkit with packaged desktop app, you can skip this step.

An Nvidia GPU is needed for faster inference (about 16ms/slice on 1080ti gpu).

Requirements:

You need to download this repository and run following command:

cd path/to/repository/
pip install -r requirement.txt

The project folder looks like this:

path
├─ ... (Codes in DABC-Net-for-COVID-19 repository. Use download.sh to get following files) 
│
├─Input_data
│      2020034797_0123_2949_20200123015940_4.nii.gz
│      2020034797_0125_3052_20200125111145_4.nii.gz
│      ...
│
├─Output_data
│   │
│   ├─covid
│   │      2020034797_0123_2949_20200123015940_4.nii.gz
│   │      2020034797_0125_3052_20200125111145_4.nii.gz
│   │      ...
│   │
│   ├─lung
│   │      2020034797_0123_2949_20200123015940_4.nii.gz
│   │      2020034797_0125_3052_20200125111145_4.nii.gz
│   │      ...
│   │
│   └─uncertainty
│           2020034797_0123_2949_20200123015940_4_predictive_aleatoric.nii.gz
│           2020034797_0125_3052_20200125111145_4_sample_1.nii.gz
│           ...
│
├─weight
│       model_05090017
│       ...
│
│ (following folders are required if you need longitudinal study)
│
├─meta
│       2020035021.csv
│
└─model
        prediction.pkl
        ...

Quick Start

DABC-Net for desktop app

Inference:

  1. Download and double click the DABC_Net.exe(Windows) or DABC_Mac(Mac OS) file. You can run our network even without installing Tensorflow or Python interpreter on you computer. The UI looks like this:

    Alt text

  2. Type or select the input folder where you store nii/nii.gz format CT scans data. The output results will be saved in the folder you specified.

  3. Choose the sform code name, the default value is ‘NIFTI_XFORM_SCANNER_ANAT’, some scans without complete header files may loss this value. For more details about header files, please see this site.

  4. Click ‘Run’ button. After all the inference done, the progress bar window will be closed.

    fig.2

    Here are some examples: fig.4

Uncertainty:

In DABC-Net, we approximate Bayersian inference using DropBlock, a form of Monte Carlo dropout. For more details about aleatory and epistemic uncertainty, please refer to this paper.

  1. Follow instructions from item(1-3) in the section above.

  2. Choose sample times (integer, e.g. 10). The network will sample 10 times to compute aleatory/epistemic uncertainty and get mean prediction outcome as final segmentation.

  3. Set ‘threshold’ to get binary output if you need. The default value is 0.5. If you want to save raw probability map from the last sigmoid activation layer of the network, just set threshold to 0.

  4. ‘Method’ denotes what kind of uncertainty you want to save.

    Here are some examples:

Visualization:

Then, choose appropriate HU range (e.g. -1024~512) via right sliding window.

DABC-Net for Colab

Open In Colab

Inference:

  1. Put your data in a folder.
  2. Select the input and output folder, and run following command: 
    DABC_infer(input_path, output_path)
    • nii_path:
    • Input: Folder path of input data(nii or nii.gz format).
    • Type: string - save_path:
    • Input: Folder path of input data(nii or nii.gz format). The segmentation results will be saved as nii.gz format.
    • Type: string - usage:
    • Input: Folder path of input data(nii or nii.gz format).
    • Type: string, ‘lung’ or ‘covid’

Uncertainty:

DABC_uncertainty(nii_filename, save_filename, sample_value, uncertainty='Aleatoric')

For more detail, please refer to notebook.

DABC-Net for Website

coming soon

Progress prediction

Model

Feature

Feature we used:

Feature Scan
Left lesion volume scan0 & scan1
Left lung volume scan0 & scan1
Left lesion ratio scan0 & scan1
Left consolidation volume scan0 & scan1
Left weighted volume scan0 & scan1
Left z-position scan0 & scan1
Right lesion volume scan0 & scan1
Right lung volume scan0 & scan1
Right lesion ratio scan0 & scan1
Right consolidation volume scan0 & scan1
Right weighted volume scan0 & scan1
Right z-position scan0 & scan1
Age scan0 & scan1
Sex scan0 & scan1

Base learner

Base learners we used:

Base learner MinMaxScaler Necessity Feature Importance
SVM True False
MLP True False
Logistic Regression True False
Naive Bayes False False
Random Forest False True
Adaboost False True
Gradient Boost False True
XGBoost False True

MinMaxScalar

For base learners sensitive to data normalization(svm, mlp, …), we provide the min max normalization based on our training dataset. The weights without min max scalar (TODO) are also provided with fewer base learners and lower performance.

Usage

Prediction

pred = predict_base_learners(base_learners, feature)

Visualization of progress

Here are some examples:

Progression curve of severe patient：

Progression curve of mild patient：

x-axis: time(day), y-axis: lesion ratio

Visualization of different time point scans

Data

Dataset with Expert Annotations and Benchmark

Data Sources

Notes

Acknowledgements: We thank COVID-19-CT-Seg-Benchmark repository for providing covid-19 segmentation dataset and benchmark. We also thank this repository for providing us ideas for designing ui.

Disclaimer: This toolkit is only for research purpose and not approved for clinical use.