COVID-19-Scanner / Git / Diff of /README.md

Models:

SCallahan/

COVID-19-Scanner

Downloads: 1

Diff of /README.md [69507b] .. [c5dfeb]

Switch to unified view


## Coronavirus disease 2019 (COVID-19) X-Ray Scanner for Diagnosis Triage
<img src="https://github.com/ajsanjoaquin/COVID-19-Scanner/blob/master/images/sample.png?raw=true">

This model is meant to help triage patients (prioritize certain patients for testing, quarantine, and medical attention)
that require diagnosis for COVID-19. This model is not meant to diagnose COVID-19.

It takes both PA and AP X-ray images (in DICOM format) as inputs and outputs a prediction for each image from one
of three labels (covid, opacity, nofinding). Those predicted with the labels "covid" and "opacity", as well as additional risk factors (Age, X-ray View Position) taken from the metadata, are flagged by the model for priority action.

The model analyzes each image for features predominant to COVID-19 images and for features present in images with Lung Opacity. If the model does not detect these features from the images, it returns a "nofinding" prediction. A "nofinding" prediction does not mean the patient is not sick for other diseases. The "covid" and "opacity" labels are not mutually exclusive. Mild and severe cases of COVID-19 exhibit lung opacity in images (which is also an idicator of Pneumonia), so the ideal worst case for the model is that an image from a patient who truly has COVID-19 symptoms is misclassified as "opacity". Therefore, the patient will still get priority action. This mitigates the current lack of COVID-19 X-ray data publicly available and the problem of false-negatives for binary classifiers.

**For example**, a 54-year old patient predicted to have lung opacity is flagged as higher priority for action by the model over a 20-year old patient predicted to have symptoms of COVID-19. Both are still flagged as higher priority for COVID-19 testing compared to those who were predicted to have no findings.

**Demo Video:**

[![Demo Video](http://img.youtube.com/vi/NSQoiGwCB80/0.jpg?raw=true)](https://www.youtube.com/watch?v=NSQoiGwCB80)

## Features

<img src="images/app%20screenshot.png">

1. Classifier (A.I. model) <br/>
2. Flagging Risk Factors <br/>
3. Visualization for Classifier explainability (Grad-CAM) <br/>

## Specifications
Architecture: Resnet 34 <br/>
Training dataset: 26,000 images with weighted resampling <br/>
Training dataset size before Resampling: <br/>

1. "covid" - 186 <br/>
2. "opacity" - 5801 <br/>
3. "nofinding" - 19884 <br/>
*Note: "nofinding" images include both healthy and non-healthy lungs that do not exhibit opacity

The model was trained on both AP and PA Chest X-rays.

Area under the Receiver Operator Characteristic (AUROC) was the chief metric used to determine model performance. It was calculated with a one-vs-all approach.

AUROC for "covid", "opacity", "nofinding" were at 99.97%, 92.64%, and 92.73%, respectively.

## Resources
Updated Model (Use with Pytorch; .pth) : https://www.dropbox.com/s/o27w0dik8hdjaab/corona_resnet34.pth?dl=0 <br/>

Datasets used: 1. https://github.com/ieee8023/covid-chestxray-dataset <br/>

2. https://www.kaggle.com/c/rsna-pneumonia-detection-challenge <br/>

Test_results: check the csv file in the repository <br/>

Colab_notebook (online version): check repository (Notice: <strong>NOT A DIAGNOSTIC TOOL</strong>) <br/>

## Installation (Application version)

I: Install [Docker](https://docs.docker.com/get-docker/) appropriate for your computer

II. In the terminal:

1. Clone the repository
2. Set the current directory to the root of this repository
3. Run ``` docker-compose up -d ``` 
4. After the setup is finished, it should open a new browser where the app is located.
5. To stop the application, run ``` docker-compose stop ```

## (Previous version) Coronavirus disease 2019 (COVID-19) X-Ray Scanner 
For the first iteration of the model, I built a CNN neural network that classifies a given Chest X-ray as positive for pneumonia caused by COVID-19 or not. This model is originally meant to demonstrate a proof-of-concept. The model was trained, and accepts, Posteroanterior views only. 

Rationale/More information: https://towardsdatascience.com/using-deep-learning-to-detect-ncov-19-from-x-ray-images-1a89701d1acd<br/>

## Acknowledgements

We thank Dr. Jianshu Weng, Mr. Najib Ninaba and their organisation, [AI Singapore](http://www.aisingapore.org/) (AISG), for their generous support in providing the infrastructure to train the latest iteration of the model.

Thank you to Ivan Leo for developing the application. Thank you to Chris, Raphael and Sunny for assisting in the application debugging.

## LICENSE
<a rel="license" href="https://opensource.org/licenses/MIT"><img alt="Creative Commons Licence" style="border-width:0" src="https://upload.wikimedia.org/wikipedia/commons/thumb/0/0c/MIT_logo.svg/220px-MIT_logo.svg.png" /></a>

	a/README.md		b/README.md
1	## Coronavirus disease 2019 (COVID-19) X-Ray Scanner for Diagnosis Triage	1	## Coronavirus disease 2019 (COVID-19) X-Ray Scanner for Diagnosis Triage
2	<img src="images/sample.png">	2	<img src="https://github.com/ajsanjoaquin/COVID-19-Scanner/blob/master/images/sample.png?raw=true">
3		3
4	This model is meant to help triage patients (prioritize certain patients for testing, quarantine, and medical attention)	4	This model is meant to help triage patients (prioritize certain patients for testing, quarantine, and medical attention)
5	that require diagnosis for COVID-19. This model is not meant to diagnose COVID-19.	5	that require diagnosis for COVID-19. This model is not meant to diagnose COVID-19.
6		6
7	It takes both PA and AP X-ray images (in DICOM format) as inputs and outputs a prediction for each image from one	7	It takes both PA and AP X-ray images (in DICOM format) as inputs and outputs a prediction for each image from one
8	of three labels (covid, opacity, nofinding). Those predicted with the labels "covid" and "opacity", as well as additional risk factors (Age, X-ray View Position) taken from the metadata, are flagged by the model for priority action.	8	of three labels (covid, opacity, nofinding). Those predicted with the labels "covid" and "opacity", as well as additional risk factors (Age, X-ray View Position) taken from the metadata, are flagged by the model for priority action.
9		9
10	The model analyzes each image for features predominant to COVID-19 images and for features present in images with Lung Opacity. If the model does not detect these features from the images, it returns a "nofinding" prediction. A "nofinding" prediction does not mean the patient is not sick for other diseases. The "covid" and "opacity" labels are not mutually exclusive. Mild and severe cases of COVID-19 exhibit lung opacity in images (which is also an idicator of Pneumonia), so the ideal worst case for the model is that an image from a patient who truly has COVID-19 symptoms is misclassified as "opacity". Therefore, the patient will still get priority action. This mitigates the current lack of COVID-19 X-ray data publicly available and the problem of false-negatives for binary classifiers.	10	The model analyzes each image for features predominant to COVID-19 images and for features present in images with Lung Opacity. If the model does not detect these features from the images, it returns a "nofinding" prediction. A "nofinding" prediction does not mean the patient is not sick for other diseases. The "covid" and "opacity" labels are not mutually exclusive. Mild and severe cases of COVID-19 exhibit lung opacity in images (which is also an idicator of Pneumonia), so the ideal worst case for the model is that an image from a patient who truly has COVID-19 symptoms is misclassified as "opacity". Therefore, the patient will still get priority action. This mitigates the current lack of COVID-19 X-ray data publicly available and the problem of false-negatives for binary classifiers.
11		11
12	For example, a 54-year old patient predicted to have lung opacity is flagged as higher priority for action by the model over a 20-year old patient predicted to have symptoms of COVID-19. Both are still flagged as higher priority for COVID-19 testing compared to those who were predicted to have no findings.	12	For example, a 54-year old patient predicted to have lung opacity is flagged as higher priority for action by the model over a 20-year old patient predicted to have symptoms of COVID-19. Both are still flagged as higher priority for COVID-19 testing compared to those who were predicted to have no findings.
13		13
14	Demo Video:	14	Demo Video:
15		15
16	[![Demo Video](http://img.youtube.com/vi/NSQoiGwCB80/0.jpg)](https://www.youtube.com/watch?v=NSQoiGwCB80)	16	[![Demo Video](http://img.youtube.com/vi/NSQoiGwCB80/0.jpg?raw=true)](https://www.youtube.com/watch?v=NSQoiGwCB80)
17		17
18	## Features	18	## Features
19		19
20	<img src="images/app%20screenshot.png">	20	<img src="images/app%20screenshot.png">
21		21
22	1. Classifier (A.I. model) <br/>	22	1. Classifier (A.I. model) <br/>
23	2. Flagging Risk Factors <br/>	23	2. Flagging Risk Factors <br/>
24	3. Visualization for Classifier explainability (Grad-CAM) <br/>	24	3. Visualization for Classifier explainability (Grad-CAM) <br/>
25		25
26	## Specifications	26	## Specifications
27	Architecture: Resnet 34 <br/>	27	Architecture: Resnet 34 <br/>
28	Training dataset: 26,000 images with weighted resampling <br/>	28	Training dataset: 26,000 images with weighted resampling <br/>
29	Training dataset size before Resampling: <br/>	29	Training dataset size before Resampling: <br/>
30		30
31	1. "covid" - 186 <br/>	31	1. "covid" - 186 <br/>
32	2. "opacity" - 5801 <br/>	32	2. "opacity" - 5801 <br/>
33	3. "nofinding" - 19884 <br/>	33	3. "nofinding" - 19884 <br/>
34	*Note: "nofinding" images include both healthy and non-healthy lungs that do not exhibit opacity	34	*Note: "nofinding" images include both healthy and non-healthy lungs that do not exhibit opacity
35		35
36	The model was trained on both AP and PA Chest X-rays.	36	The model was trained on both AP and PA Chest X-rays.
37		37
38	Area under the Receiver Operator Characteristic (AUROC) was the chief metric used to determine model performance. It was calculated with a one-vs-all approach.	38	Area under the Receiver Operator Characteristic (AUROC) was the chief metric used to determine model performance. It was calculated with a one-vs-all approach.
39		39
40	AUROC for "covid", "opacity", "nofinding" were at 99.97%, 92.64%, and 92.73%, respectively.	40	AUROC for "covid", "opacity", "nofinding" were at 99.97%, 92.64%, and 92.73%, respectively.
41		41
42	## Resources	42	## Resources
43	Updated Model (Use with Pytorch; .pth) : https://www.dropbox.com/s/o27w0dik8hdjaab/corona_resnet34.pth?dl=0 <br/>	43	Updated Model (Use with Pytorch; .pth) : https://www.dropbox.com/s/o27w0dik8hdjaab/corona_resnet34.pth?dl=0 <br/>
44		44
45	Datasets used: 1. https://github.com/ieee8023/covid-chestxray-dataset <br/>	45	Datasets used: 1. https://github.com/ieee8023/covid-chestxray-dataset <br/>
46		46
47	2. https://www.kaggle.com/c/rsna-pneumonia-detection-challenge <br/>	47	2. https://www.kaggle.com/c/rsna-pneumonia-detection-challenge <br/>
48		48
49	Test_results: check the csv file in the repository <br/>	49	Test_results: check the csv file in the repository <br/>
50		50
51	Colab_notebook (online version): check repository (Notice: <strong>NOT A DIAGNOSTIC TOOL</strong>) <br/>	51	Colab_notebook (online version): check repository (Notice: <strong>NOT A DIAGNOSTIC TOOL</strong>) <br/>
52		52
53	## Installation (Application version)	53	## Installation (Application version)
54		54
55	I: Install [Docker](https://docs.docker.com/get-docker/) appropriate for your computer	55	I: Install [Docker](https://docs.docker.com/get-docker/) appropriate for your computer
56		56
57	II. In the terminal:	57	II. In the terminal:
58		58
59	1. Clone the repository	59	1. Clone the repository
60	2. Set the current directory to the root of this repository	60	2. Set the current directory to the root of this repository
61	3. Run ``` docker-compose up -d ```	61	3. Run ``` docker-compose up -d ```
62	4. After the setup is finished, it should open a new browser where the app is located.	62	4. After the setup is finished, it should open a new browser where the app is located.
63	5. To stop the application, run ``` docker-compose stop ```	63	5. To stop the application, run ``` docker-compose stop ```
64		64
65	## (Previous version) Coronavirus disease 2019 (COVID-19) X-Ray Scanner	65	## (Previous version) Coronavirus disease 2019 (COVID-19) X-Ray Scanner
66	For the first iteration of the model, I built a CNN neural network that classifies a given Chest X-ray as positive for pneumonia caused by COVID-19 or not. This model is originally meant to demonstrate a proof-of-concept. The model was trained, and accepts, Posteroanterior views only.	66	For the first iteration of the model, I built a CNN neural network that classifies a given Chest X-ray as positive for pneumonia caused by COVID-19 or not. This model is originally meant to demonstrate a proof-of-concept. The model was trained, and accepts, Posteroanterior views only.
67		67
68	Rationale/More information: https://towardsdatascience.com/using-deep-learning-to-detect-ncov-19-from-x-ray-images-1a89701d1acd<br/>	68	Rationale/More information: https://towardsdatascience.com/using-deep-learning-to-detect-ncov-19-from-x-ray-images-1a89701d1acd<br/>
69		69
70	## Acknowledgements	70	## Acknowledgements
71		71
72	We thank Dr. Jianshu Weng, Mr. Najib Ninaba and their organisation, [AI Singapore](http://www.aisingapore.org/) (AISG), for their generous support in providing the infrastructure to train the latest iteration of the model.	72	We thank Dr. Jianshu Weng, Mr. Najib Ninaba and their organisation, [AI Singapore](http://www.aisingapore.org/) (AISG), for their generous support in providing the infrastructure to train the latest iteration of the model.
73		73
74	Thank you to Ivan Leo for developing the application. Thank you to Chris, Raphael and Sunny for assisting in the application debugging.	74	Thank you to Ivan Leo for developing the application. Thank you to Chris, Raphael and Sunny for assisting in the application debugging.
75		75
76	## LICENSE	76	## LICENSE
77	<a rel="license" href="https://opensource.org/licenses/MIT"><img alt="Creative Commons Licence" style="border-width:0" src="https://upload.wikimedia.org/wikipedia/commons/thumb/0/0c/MIT_logo.svg/220px-MIT_logo.svg.png" /></a>	77	<a rel="license" href="https://opensource.org/licenses/MIT"><img alt="Creative Commons Licence" style="border-width:0" src="https://upload.wikimedia.org/wikipedia/commons/thumb/0/0c/MIT_logo.svg/220px-MIT_logo.svg.png" /></a>