BrixIA COVID-19 Project

What do you find here

Info, code (BS-Net), link to data (BrixIA COVID-19 Dataset annotated with Brixia-score), and additional material related to the BrixIA COVID-19 Project

Defs

BrixIA COVID-19 Project: go to the webpage Brixia score: a multi-regional score for Chest X-ray (CXR) conveying the degree of lung compromise in COVID-19 patients

BS-Net: an end-to-end multi-network learning architecture for semiquantitative rating of COVID-19 severity on Chest X-rays

BrixIA COVID-19 Dataset: 4703 CXRs of COVID-19 patients (anonymized) in DICOM format with manually annotated Brixia score

Project paper

Preprint avaible here

@article{SIGNORONI2021102046,
title = {BS-Net: learning COVID-19 pneumonia severity on a large Chest X-Ray dataset},
journal = {Medical Image Analysis},
pages = {102046},
year = {2021},
issn = {1361-8415},
doi = {https://doi.org/10.1016/j.media.2021.102046},
url = {https://www.sciencedirect.com/science/article/pii/S136184152100092X},
author = {Alberto Signoroni and Mattia Savardi and Sergio Benini and Nicola Adami and Riccardo Leonardi and Paolo Gibellini and Filippo Vaccher and Marco Ravanelli and Andrea Borghesi and Roberto Maroldi and Davide Farina},
}2020}
}

Overall Scheme

Table of Contents

• Data
• Getting Started
• License
• Citation

Datasets

BrixIA COVID-19 Dataset

The access and use, for research purposes only, of the annotated BrixIA COVID-19 CXR Dataset have been granted form the Ethical Committee of Brescia (Italy) NP4121 (last update 08/07/2020).

The data can be downloaded from the website https://brixia.github.io/.

To unpack all the zipped archives, on unix-like system do:

1. Download all the files
2. From the command line call: cat *.tar.gz.* | tar -xzv
3. A folder called dicom_clean will be created with all the unpacked files

Instead, for MS Window: 2. type *.tar.gz.* | tar xvfz -

[Update] We revised the dataset and removed the DICOM found to have acquisition problems (low quality). The total now is 4695.

Annotation and CXR from Cohen's dataset

We exploit the public repository by Cohen et al. which contains CXR images (We downloaded a copy on May 11th, 2020).

In order to contribute to such public dataset, two expert radiologists, a board-certified staff member and a trainee with 22 and 2 years of experience respectively, produced the related Brixia-score annotations for CXR in this collection, exploiting labelbox, an online solution for labelling. After discarding problematic cases (e.g., images with a significant portion missing, too small resolution, the impossibility of scoring for external reasons, etc.), the final dataset is composed of 192 CXR, completely annotated according to the Brixia-score system.

Below a list of each field in the annotation csv, with explanations where relevant

Scheme

Attribute	Description
filename	filename from Cohen dataset
from S-A to S-F	The 6 regions annotatated by a Senior radiologist (+20yr expertise), from 0 to 3
S-Global	Global score by the Senior radiologist (sum of S-A : S-F), from 0 to 18
from J-A to J-F	The 6 regions annotatated by a Junior radiologist (+2yr expertise), from 0 to 3
J-Global	Global score by the Junior radiologist (sum of S-A : S-F), from 0 to 18

Segmentation Dataset

We provide the script to prepare the dataset as described in the Project paper.

We exploit different segmentation datasets in order to pre-train the extended-Unet module of the proposed architecture. We used the original training/test set splitting when present (as the case of the JSRT database), otherwise we took the first 50 images as test set, and the remaining as training set (see Table below).

Table

	Training-set	Test-set	Split
Montgomery County	88	50	first 50
Shenzhen Hospital	516	50	first 50
JSRT database	124	123	original
------	-----	-----	-----
Total	728	223

The data can be downloaded from their respective sites.

Alignment synthetic dataset

To avoid the inclusion of anatomical parts not belonging to the lungs in the AI pipeline, which would increase the task complexity or introduce unwanted biases, we integrated into the pipeline an alignment block. This exploits a synthetic dataset (used for on-line augmentation) composed of artificially transformed images from the segmentation dataset (see Table below), including random rotations, shifts, and zooms, which is used in the pre-training phase.

The parameters refer to the implementation in Albumentation. In the last column is expressed the probability of the specific transformation being applied.

Additional details

	Parameters (up to)	Probability
Rotation	25 degree	0.8
Scale	10%	0.8
Shift	10%	0.8
Elastic transformation	alpha=60, sigma=12	0.2
Grid distortion	steps=5, limit=0.3	0.2
Optical distortion	distort=0.2, shift=0.05	0.2

Getting Started

Install Dependencies

The provided code is written for Python 3.x. To install the needed requirements run:

pip install -r requirements.txt

For the sake of performance, we suggest to install tensorflow-gpu in place of the standard CPU version.

Include the src folder in your python library path or launch python from that folder.

Load Cohen dataset with BrixiaScore annotations

from datasets import brixiascore_cohen  as bsc

# Check the docsting for additional info
X_train, X_test, y_train, y_test = bsc.get_data()

Prepare and load the segmentation dataset

To prepare the segmentation dataset either Montgomery County, Shenzhen Hospital, and JSRT datasets must be downloaded from their websites and unpacked in a folder (for instance data/sources/). Than execute:

 python3 -m datasets.lung_segmentation  --input_folder data/sources/ --target_size 512

or just import it (the first time it is executed, it will create the segmentation dataset):

from datasets import lung_segmentation  as ls

# Check the docsting for additional info. The train-set is provided as a generator, while the validation set is
# preloaded in memory.
# `get_data` accepts a configuration dictionary where you can specify every parameter. See `ls.default_config`
train_gen, (val_imgs, val_masks) = ls.get_data()

Prepare and load the alignment dataset

To prepare the alignment dataset, the segmentation one mush be already built (see previous point)

from datasets import synthetic_alignment  as sa

# Check the docsting for additional info. The train-set and validation-set are provided as generators
# `get_data` accepts a configuration dictionary where you can specify every parameter. See `ls.default_config`
train_gen, val_gen = sa.get_data()

Model weights

The model weight and a demo notebook can be found here

Other steps

Instructions for preparing and loading the Brixia Covid-19 Dataset and the BS-Net will follow (see specific sections for more info).

License and Attribution

Disclaimer

The BS-Net model and source code, the BrixIA COVID-19 Dataset, and the Brixia score annotations, are provided "as-is" without any guarantee of correct functionality or guarantee of quality. No formal support for this software will be given to users. It is possible to report issues on GitHub though. This repository and any other part of the BrixIA COVID-19 Project should not be used for medical purposes. In particular this software should not be used to make, support, gain evidence on and aid medical decisions, interventions or diagnoses. Specific terms of use are indicated for each part of the project.

Data

• BrixIA COVID-19 dataset: access conditions and term of use are reported on the dataset website.
• Pulic Cohen dataset: Each image has license specified in the original file by Cohen's repository file. Including Apache 2.0, CC BY-NC-SA 4.0, CC BY 4.0. There are additional 7 images from Brescia under a CC BY-NC-SA 4.0 license.
• Brixia-score annotations for the pulic Cohen's dataset are released under a CC BY-NC-SA 4.0 license.

Code

• Released under Open Source license.

Contacts

Alberto Signoroni [email protected]

Mattia Savardi [email protected]

Citations

For any use or reference to this project please cite the following paper.

[NEWS]: this work got accepted at Medical Image Analysis. Available here

@article{SIGNORONI2021102046,
title = {BS-Net: learning COVID-19 pneumonia severity on a large Chest X-Ray dataset},
journal = {Medical Image Analysis},
pages = {102046},
year = {2021},
issn = {1361-8415},
doi = {https://doi.org/10.1016/j.media.2021.102046},
url = {https://www.sciencedirect.com/science/article/pii/S136184152100092X},
author = {Alberto Signoroni and Mattia Savardi and Sergio Benini and Nicola Adami and Riccardo Leonardi and Paolo Gibellini and Filippo Vaccher and Marco Ravanelli and Andrea Borghesi and Roberto Maroldi and Davide Farina},
}