How to contribute?

[arthurfigueiredo]

Hi,

I'm also a UFSC Araranguá student from course of Information Tecnology and Comunication.

I've experience working with python, data structures, algorithms, database etc.

I'm looking for a completion course work and I'm interested in this problem, how can i contribute to this project?

my email is arthurfn.webdev@gmail.com

[matheushent]

Hello Arthur,

Very glad of your interest. Right now I'm working on developing some type of defence against adversarial attacks. I think that small changes in x-ray images or tomographic image can indicate if someone's COVID or doesn't.

It would be very good if you could start contributing from this point. Another good point is to get more images.

It would be very good if you try models using 3 types of classification: presence of no virus, presence of any virus but COVID and presence of COVID. For this we need x-ray images of lungs with another type of virus. Fortunately, we can get those images here.

Feel free to contact me: matheustosta66@gmail.com +5566996655998

[arthurfigueiredo]

I created a repository to gather chest x-ray and CT images, the goal is to create a collection that can be useful for other projects that are analyzing the covid-19 with computer vision.

https://github.com/arthurfigueiredo/covid-dataset/

[matheushent]

I created a repository to gather chest x-ray and CT images, the goal is to create a collection that can be useful for other projects that are analyzing the covid-19 with computer vision.

https://github.com/arthurfigueiredo/covid-dataset/

Very good organization; it will be very useful. Now I think we can start testing training with the 3 categories I said above.

[erikson84]

Great initiative! I was working on the available dataset using pytorch and fastai libraries and I am getting very good accuracy with resnet32, very quickly (around 3-5 epochs) using the fit1cycle regimen and freezing the NN main layers (the small sample size causes great variation in out-of-sample error estimates, but around 85%-90% accuracy).

I talked about it with a friend of mine who is a doctor (who specializes in intensive care and pneumonology), and he said that x-ray on a patient who is lying down has low sensitivity; and a CT scan is too costly to be applied in large scale. He mentioned ultrasound as a quick, cheap and accurate alternative in practice. I currently looking for ultrasound images libraries, but they are not as easy to find as x-ray or CT scans.

I will update here as soon as I find something.

[matheushent]

Great initiative! I was working on the available dataset using pytorch and fastai libraries and I am getting very good accuracy with resnet32, very quickly (around 3-5 epochs) using the fit1cycle regimen and freezing the NN main layers (the small sample size causes great variation in out-of-sample error estimates, but around 85%-90% accuracy).

I talked about it with a friend of mine who is a doctor (who specializes in intensive care and pneumonology), and he said that x-ray on a patient who is lying down has low sensitivity; and a CT scan is too costly to be applied in large scale. He mentioned ultrasound as a quick, cheap and accurate alternative in practice. I currently looking for ultrasound images libraries, but they are not as easy to find as x-ray or CT scans.

I will update here as soon as I find something.

Great idea; I will look deeply in using ultrasound images.

Yesterday I talked with a doctor from Florianópolis (my city), a specialist in radiology, he said the usage of x-ray images is very plausible and can be used as a trial, selecting the most serious cases by order of priority. Anyway, I guess we can keep making tests with different type of images.

Any update about dataset with ultrasound images please contact us.

[rlnovak]

Hi! This is a good initiative. Last year I have worked on a project involving neural networks for object identification and got interesting results with transfer learning on Inception_v3 and Xception convnets. I would like to try with the covid19 chest x-ray images.

Just two things I would like to ask:

1) In the https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia repository, the chest x-rays were taken from children, as the repo owner points out:

"Chest X-ray images (anterior-posterior) were selected from retrospective cohorts of paediatric patients of one to five years old from Guangzhou Women and Children’s Medical Center, Guangzhou. All chest X-ray imaging was performed as part of patients’ routine clinical care."

I don't know if this could generate a bias when the vast majority of the covid19 chest x-rays are from adults. A doctor could answer this question.

2) There are both chest x-rays and CT scans available in the repositories. These types of images are different. Are you training the networks with x-rays and CT scans mixed up? I wonder if this could create problems, because of the different features these images will generate.

Finally, it is interesting to take a look at this: https://arxiv.org/abs/2003.09871. They are using a very complex resnet with depthwise separable convolutions, and the same repositories we are using. As we can see, the big problem is still the availability of images (as usual in projects involving convnets and image recognition).

[matheushent]

Hi! This is a good initiative. Last year I have worked on a project involving neural networks for object identification and got interesting results with transfer learning on Inception_v3 and Xception convnets. I would like to try with the covid19 chest x-ray images.

Just two things I would like to ask:

1. In the https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia repository, the chest x-rays were taken from children, as the repo owner points out:

"Chest X-ray images (anterior-posterior) were selected from retrospective cohorts of paediatric patients of one to five years old from Guangzhou Women and Children’s Medical Center, Guangzhou. All chest X-ray imaging was performed as part of patients’ routine clinical care."

I don't know if this could generate a bias when the vast majority of the covid19 chest x-rays are from adults. A doctor could answer this question.

1. There are both chest x-rays and CT scans available in the repositories. These types of images are different. Are you training the networks with x-rays and CT scans mixed up? I wonder if this could create problems, because of the different features these images will generate.

Finally, it is interesting to take a look at this: https://arxiv.org/abs/2003.09871. They are using a very complex resnet with depthwise separable convolutions, and the same repositories we are using. As we can see, the big problem is still the availability of images (as usual in projects involving convnets and image recognition).

About point 2: Yes, I'm training with x-rays and CT scans mixed up. I thought a lot about it and I tried without mix and with mix. My decision of keeping mixed on repo is totally arbitrary. I think the best way to have good results is to keep the neural network focused on just one type of image.

Thank you for pointing out this great article. Any update feel free to contribute here.