Deep Learning in Medicine resources & Study Group

In the spirit of the now extremely successful Geospatial Deep Learning resources & study group started by @daveluo and Time series/ sequential data study group started by @oguiza, this discussion thread is dedicated as a resource sharing and gathering space for people enthusiastic about applying deep learning techniques and what we have learned in Part 1 and will learn in Part 2 to solve the challenges we face in medicine and healthcare.

To get things started, here is a list of resources and recent relevant posts. This post should be made a wiki soon and you will be able to add your content in no time.

If any of you are enthusiastic about healthcare and want to focus on related applications when taking Part 1/2, please leave a reply below and tell us what you want to work on. It would be super exciting to work with a group of fellow fastai students to understand, implement, and possibly beat the papers in this field!


Work by Fellow Fastai Students

Please kindly put your work here! Remember to @ yourself so people can know who you are. :wink:


You could always find more here in the medicine category of Papers with Code.




  • Stanford Machine Learning Group
    • Led by Andrew Ng
    • Though with a very general name, it focuses very specifically on healthcare.
    • Hosted some very interesting medical image challenges with very detailed and user-friendly guide on how to participate.
  • Applied Deep Learning in Radiology, Oncology and Pathology
    • A series of conference tutorials that use PyTorch and deep learning techniques to solve classification, segmentation, and object detection tasks in medicine.
    • Discord study/working group for Fast.AI people interested in medical collaboration, paper discussion, and problem solving.

Note that this is a forum wiki thread, so you all can edit this post to add/change/organize info to help make it better! To edit, click on the little edit icon at the bottom of this post. Here’s a pic of what to look for:


@sgugger Hey dear, could you kindly use your superpower to wikify it? Many thanks.

Meanwhile @PegasusWithoutWinds do you mind adding the resources that have been mentioned in This little interview with Dr. Alexandre Cadrin-Chenevert that I had the honour of doing?

Alexandre has shared many cool resources in the interview.
I would also suggest that we add relavant kaggle competitions since sometimes experts like Alexandre would make really cool kernels sharing their domain knowledge.

1 Like

Oh absolutely. The interview looks wonderful. Thanks again!

1 Like

Very good initiative. I have worked with the DDSM mammography dataset i part 1

1 Like

Hey @Kaspar, great to have you.

Please add the dataset and your previous work in the top post once it gets wikified.

Thank you for setting up this thread!

Some time ago I implemented the starter notebook from “A primer on deep learning in genomics” in plain pytorch and fastai:

I’m very interested in health and biotech deep learning applications and I’m looking forward to the discussions in this thread! :slight_smile:


Hey @MicPie, those notebooks look wonderful! Would you mind putting them under “Work by Fellow Fastai Students” in the top post?

1 Like

Hello @PegasusWithoutWinds. Thank you for this initiative! I think like you that DL, and fastai in particular, can help a lot of people (sick or not) with respect to medicine.

To avoid loosing information or repeating links already posted, we could create a shared Google Drive. It could additionally be used to store datasets. What do you think?

1 Like

Oh please feel free to do so if you have enough Google Drive storage. We can then put the link in the top post.

I just added the link to the ACR-DSI TOUCH-AI use cases under the Applications heading in the wiki. This is a list of use cases that have been vetted and determined to be of significant potential value by radiologists and other experts in medical imaging. Something to consider when thinking about potential projects…


Thanks for initiating this thread I am also interested in deep learning in the medicine space.

This is wonderful! Thank you Walter! Domain expert’s opinions are gold.

We data scientists only solve problems. You guys determine what are important and worthy of the effort!

PS: I just moved it to the overview section.

Hey @jcreinhold @zearo, would you please kindly put your amazing work under “Work by Fellow Fastai Students” and write a short summary for it? I am sure that many other fellow fastai students will benefit from them.

Hey, thank you for setting up this thread!

I would like to ingest a series of 2d tiff images (each folder is a single scan) as a 3d image in order to pass it through a 3d CNN resnet in order to perform classification and segmentation of the scan.

I was looking through the source code of Fastai scans by @renato and read through the blog post by @jcreinhold and saw each used ItemLists extensively.

What is the best way to get this done? This notebook seems like a good start.

How with Fastai scans am I supposed to save the data in my filesystem for proper loading?

1 Like

You should be able to read each tiff image into PyTorch and then concatenate them together into a 3D tensor.



Have been doing some research and found a pretty cool resource for 3d segmentation which I believe is based on the winning solution for thought worth sharing:


Hello Everybody,

Thank you for creating this group as medical image analysis is really an interesting and deep topic. Being a newbie in this domain I am still exploring and learning things. I came across this topic of Mitosis count problem in medical images and found it really very interesting.

I would love to dive deep in this, so asking some suggestions here - where should I go for content in getting some domain knowledge and also maybe to get a head start a solution solving this mitosis count problem.


Checkout similar dataset and solutions, possibly in Kaggle. After identifying the primary concepts and techniques involved, try to learn about them one at a time very well.



I have a basic model for medical imaging classification. I’m looking to start a pilot on a medical site for research.

What are the typical steps needed (i.e. de-identification)?

What are the advantages/disadvantages of having the model served on site vs in the cloud?

Are there any guides/tutorials/open source tools which are recommended?