CT-Angel

---

Table of Contents

• Background
• Data
• Model Development
• Usage
• Directory Structure
• About the author

---

1. Background

Computed tomography (CT) is the preferred imaging method for diagnosing 2019 novel coronavirus (COVID19) pneumonia. CT-Angel aimed to construct a system based on deep learning for detecting COVID-19 pneumonia on high resolution CT, relieve working pressure of radiologists and contribute to the control of the epidemic.

2. Datasets

2.1 Datasets source

• Hospital
  • From Renmin Hospital of Wuhan University (Wuhan, Hubei province, China).
• Instruments
  • Included Optima CT680, Revolution CT and Bright Speed CT scanner (all GE Healthcare).

2.2 Datasets amount

2.2.1 Raw datasets

• 46,096 CT images from 106 patients
  • 25989 images from 51 COVID19 patients
  • 20107 images from 55 controls

2.2.2 Selected datasets(after filtering those images without good pneumonia)

• 35355 CT images selected
  • 20886 images from 51 COVID19 patients
  • 14469 images from 55 controls

3. Model development

3.1 Workflow diagram

3.2 Model training

3.2.1 Training the model for extracting valid regions of CT images

We first trained UNet++ to extract valid areas in CT images using 289 randomly selected CT images and tested it in other 600 randomly selected CT images. The training images were labelled with the smallest rectangle containing all valid areas by researchers. The model successfully extracted valid areas in 600 images in testing set with an accuracy of 100%.

3.2.2 Training the model for detecting suspicious lesions in CT images

For detecting suspicious lesions on CT scans, 691 images of COVID-19 pneumonia infection lesions labelled by radiologists and 300 images randomly selected from patients of non-COVID-19 pneumonia were used. Taking the raw CT scan images as input with a resolution of 512×512, and the labelled map from the expert as output, UNet++ was used to train in Keras in an image-to-image manner. The suspicious region was predicted under a confidence cutoff value of 0.50, and a prediction box pixel of over 25.

3.3 Model testing

3.3.1 Testing datasets

Methods	COVID-19 Patients amount	COVID-19 images amount	Other Patients amount	Other images amount
Retrospective testing	11	4382	31	9369
Prospective testing	16	/	11	/

3.3.2 Testing results

Methods	Sensitivity	Specificity	Accuracy	PPV	NPV
Retrospective testing
Per patient	100%	93.55%	95.24%	84.62%	100%
Per image	94.34%	99.16%	98.85%	88.37%	99.61%
Prospective testing(Per patient)	100%	81.82%	92.59%	88.89%	100%

Ps: PPV: positive prediction value,NPV: negative prediction value.

3.4 Representative images of model testing

• A. Computed tomography (CT) images of COVID19 pneumonia. The predictions between the artificial intelligence model and radiologists were consistent. Green boxes, labels from radiologists; red boxes, labels from the model.
• B. CT images of the control.

4. Usage

4.1 Environmental dependence

4.1.1 Computer reference configuration

Item	Configuration
Operating system	Microsoft Windows 10 64-bit
CPU	Intel Corei7-8700 3.2GHz
RAM	>=16G
Hard disk	256G solid state hard disk + 1T mechanical hard disk
Graphics card	NVIDIA GeForce RTX 2080 (8G video memory)

4.1.2 Dependent tools

Tools	Version
Python	3.5+
Tensorflow	1.13.1
Keras	2.2.4
Unet++	/
OpenCV	3.4.2
CUDA	10.0.130
CUDnn	7.6.0

4.2 Instructions for use

4.2.1 Ready to adapt the hardware

4.2.2 Download and install the required toolkit

• Python 3.5+ https://www.anaconda.com/download/

   It is recommended to download the anaconda3 integrated environment directly, which is more convenient.

• Tensorflow 1.13.1

  We chose to install TensorFlow-GPU on Anaconda:

  1. Create an environment dedicated to TensorFlow：

     conda create -n tf python=3.6
     

  2. Then enter this environment. In order to ensure that there are no problems with the installation, it is recommended to update the pip and setuptool tools：

     activate tf #Enter the tensorflow environment
     deactivate #Exit the tensorflow environment
     python -m pip install --upgrade pip 
     pip install -–upgrade setuptools
     

  3. Then you can install TensorFlow:

     pip install tensorflow-gpu==1.13.1
     

  4. After the installation is complete, enter the python interpreter and import TensorFlow. If the import is successful, the installation is successful.

• Keras 2.2.4

  pip install Keras==2.2.4

• Unet++

  We have prepared a package for Unet ++ for you, which is the folder named unet_pp.

• OpenCV 3.4.2

  pip install opencv-python==3.4.2

• CUDA 10.0.130 https://developer.nvidia.com/cuda-10.0-download-archive
• CUDnn 7.6.0 https://developer.nvidia.com/rdp/cudnn-download

  Need to register, install after registration.After decompression, place the three files in the corresponding directories of cuda.

4.2.3 Download our trained model

We open sourced two models we trained: the model for extracting valid regions of CT images and the model for detecting suspicious lesions in CT images. You can use 'ct_angel_lung_weights' to extract the valid regions of the CT images, and then use 'ct_angel_lesion_weights' to predict the location of the lesion in the extracted valid regions, which can reduce false positives.Here is our usage example:

Here are the download links for different ways, you can click any of them to download according to your network situation:

_______________________________________________________________________
Baidu cloud disk link：https://pan.baidu.com/s/1OWxNBVvL3kINV1Lf_lrdlA
extraction code：q916
———————————————————————————————————————————————————————————————————————
Google drive link: https://drive.google.com/open?id=1bWmmCroyXs0VxapeVukvtR5pTlgeOEjh
_______________________________________________________________________
Dropbox link: https://www.dropbox.com/s/7l6ocnurn4f1bht/ct-angel-weights.zip?dl=0
_______________________________________________________________________

4.2.4 Download our project

4.2.5 Run our project locally

Two jupyter files are provided in our project: one named dcm_2_jpg and the other named ct_angel_train. If your CT data is in 16-bit dicom format, you can use the code in dcm_2_jpg.ipynb to convert them to 8-bit jpg. After that, you can follow the prompts of the code comments in ct_angel_train.ipynb to place the training images and corresponding mask images in the corresponding path and folder, and then run the jupyter file to start our project and get the prediction result of your own CT images.

5. Directory Structure

├── Readme.md                            // ct-angel introduction&help
└── ct-angel-train                       // ct-angel use package
│   ├── data                             // ct-angel data
│      └── model_data                    // ct-angel model data
│          └── ct-angel_weights.txt      // model download link file
│      └── train_images                  // ct-angel model train_images
│          └── DJ20200112A0189(243).jpg  // train_images sample 
│      └── train_labels                  // ct-angel model train_labels 
│          └── DJ20200112A0189(243).jpg  // train_labels sample
│      └── val_images                    // ct-angel model val_images 
│          └── DJ20200112A0189(243).jpg  // val_images sample 
│      └── val_labels                    // ct-angel model val_labels 
│          └── DJ20200112A0189(243).jpg  // val_labels sample 
│      └── test_images                   // ct-angel model test_images
│          └── DJ20200112A0189(243).jpg  // test_images sample
│      └── test_labels                   // ct-angel model test_labels
│          └── DJ20200112A0189(243).jpg  // test_labels sample
│   ├── sample                           // ct-angel prediction sample
│      └── sample01.jpg                  // sample image
│   ├── unet_pp                          // unet++ package
│   ├── ct_angel_train.ipynb             // ct-angel-train-code
│   └── dcm_2_jpg.ipynb                  // dicom-2-jpg-code

6. About the author

• Author
  • Renmin Hospital of Wuhan University & Wuhan EndoAngel Medical Technology Co., Ltd.
• E-mail
  • [email protected]
  • [email protected]
  • [email protected]