Are you a new(ish) Graduate or a super-enthusiastic Undergraduate student working with Medical Image data, and pondering over how to get that Deep Learning model to train with it? If so, this repository is for you:
It supports the BENDER (BEst practices in medical imagiNg DEep leaRning) series of videos, which is submitted to the MICCAI Education Challenge, 2022.
We invite you to watch along the following four episodes (click on the pictures below to watch on YouTube) that track the life and times of a new student who has just started working with this kind of data: through the ups and downs of the journey to build a State-of-The-Art model.
We hope through the experience of this student, you learn not to make the same mistakes, and get a head-start in your own exploration.
See checklist for a list of keepawakes while dealing with clinical data. We believe checking off all of these as a bare minimum would help avoid 💣 surprises later on.
A downloadable PDF version is here.
Exploring the Dermamnist data set is a notebook that uses data from MedMNIST to demonstrate examples of the kinds of details you should look for, while analyzing your clinical data.
If you are lucky (or not 😉) to be working with DICOM data, Exploring DICOM tags is a notebook to demonstrate how to go about looking at the non-image metadata as well!
Here is a wonderful (non-medical-imaging) example of how simply "looking" at the data prior to modeling helps make sense of the landscape, and is a step that one must not neglect while starting off!
For more specialized data, there are standards like BIDS, Brain Imaging Data structure, for organizing multiple subject files for neuroimaging. Consider conforming to such formats to ensure easier read/write/convert workflows with your data set.
In this context, Data-centered AI is gaining more attention, especially through these papers:
- DataPerf: Benchmarks for Data-Centric AI Development
- Advances, challenges and opportunities in creating data for trustworthy AI
- IEEE Spectrum article: Andrew Ng, AI Minimalist: The Machine-Learning Pioneer Says Small is the New Big
Click here for a glossary of common terms used in the Deep Learning world, with a special focus on Medical Imaging and clinical lingo.
A downloadable PDF version is here.
Follow along this episode for the next step of actually building a model.
We start with a naive implementation (as ipynb or a py script), walking through the process of improving it iteratively in seven versions to finally beat the benchmark listed on the MedMNIST webpage in an organized experimental fashion!
In this final episode, the focus is on evaluation and deployment: specific points that are important to keep in mind for clinical relevance, out-of-distribution data, and other interesting bits.
We hope you learn something new about how to get started with Medical Imaging and Deep Learning, and more importantly, that you have fun while learning! (If you have suggestions for improvement, please do not hesitate to create an issue here)
👋 The Medical Imaging Analysis group at Universität Bern