Stanford MURA (X-Ray) Classification Competition

Have you/anyone in this thread figured out a way to leverage fast.ai image data bunch API to this sort of multi-image classification dataset? What I’m looking for is a natural workflow where we can make use of all the data augmentation functionality (I’m less worried about the visualization bit) but on multiple images that are “stuck” together to a label. E.g one augmentation would be a 4 degree rotation but the same augmentation is applied to all the same images…? This would be very helpful, any thoughts?

1 Like

There are a few things that immediately come into my head:

  1. There is a specialized item class customized for the multi-item dataset. Check out this tutorial.

  2. It is not necessary to apply the same kind of data augmentation to all the images in a study.

This is perfect! Thanks so much : ) I will keep you posted on how this turns out.

1 Like

I found something:
http://dmery.sitios.ing.uc.cl/Prints/ISI-Journals/2015-JNDE-GDXray.pdf

That is nice! Even though it is not medical X-Ray, it is still X-Ray!

As a radiologist, I do not think that having the same data augmentation (at least for rotation and magnification) is very important, x-rays, even the same set can be shot with different positions, a few degree of rotations are just due to positioning. Some difference in magnification is also present depending on the distance of the area of interest from the detector.

4 Likes

Also shear for varying degrees of compression in mammography.

1 Like

Hi. Part 2 of my journey in Deep Learning for medical images with the fastai framework on the MURA dataset.

I got a better kappa score but I need radiologists to go even further (and fastai specialists too :slight_smile: ). I guess this the right place for asking :slight_smile:

Please, feel free to use (and improve) my notebook (ensemble models, squeezenet models, etc.).

2 Likes

Great work, @pierreguillou! Definitely check out the Discord group. There are lots of great discussions going on and we have quite a few domain experts there.

Great work, I looked at your notebook, the heatmap of the “top losses” images is often outside the body part being imaged. It may be helpful to preprocess the images and remove the background, it helped me in the Kaggle Humpback Whale Identification competition

1 Like

Thanks @PegasusWithoutWinds. I just entered the Discord Group and gave you my kaggle username in order to join https://www.kaggle.com/fastaimed

Just one question: why not using only this thread?

Hi @agentili. Thank for your message. I agree with your observation.

In general, I do not like preprocessing images. The risk is too great to alter the information carried by the images.
But I like testing new things. Could you share your code?

Hi to all. I published my notebook about the MURA classification.
Could you publish yours as well in order to share ideas and code? Thanks.

I think that would be preferable if it’s an option - splitting the conversation over multiple channels is rather awkward.

1 Like

The Discord group is mostly for group video chat.

1 Like

Let me clean it up a bit and then I will put in the Kaggle Kernel. I should have done it but other things always pop up. :pensive:

Guys, I just found an excellent paper that contains the solution most pertinent to the problem we have at hand with the MURA dataset. Here is the link:

For those of you who have carefully inspected the dataset, the picture below alone should be enough to get you excited:

3 Likes

In an x-ray, the border around the image, does not contain any useful information, it is only part of the detector that is not exposed to the x-rays. The cropping code is not my code. I got it from https://www.kaggle.com/martinpiotte/bounding-box-model/output and https://www.kaggle.com/iafoss/similarity-densenet121-0-805lb-kernel-time-limit

If it is ok I would be interested in joining the collaboration on this task.
I started on this project when it was launched with the previous fastai library. These kins of challenges are what motivates me most regarding AI. To make things that help others.

2 Likes

Oh please join us. There are still tons of problems to fix.

1 Like