Today, we're excited to announce our first research competition!
INTRODUCING NANOPATH: a framework and challenge to train the best pathology foundation model in just 1 hour!
A quick thread on why we made this challenge and how to participate!
We're launching a new research competition!
Train the best pathology foundation model in just 1 hour!
The competition has already been ongoing in @MedARC_AI community for a few weeks now and we've already been making a lot of great progress!
Learn how to participate in the thread below...
Today, we're excited to announce our first research competition!
INTRODUCING NANOPATH: a framework and challenge to train the best pathology foundation model in just 1 hour!
A quick thread on why we made this challenge and how to participate!
We are supporting a new research competition in our @MedARC_AI open-source research competition. This is highly impactful, pathology foundation models can help improve diagnosis and treatment of a variety of diseases, especially cancer. We hope this competition will accelerate innovation in the space!
Learn how to participate! ↓
Today, we're excited to announce our first research competition!
INTRODUCING NANOPATH: a framework and challenge to train the best pathology foundation model in just 1 hour!
A quick thread on why we made this challenge and how to participate!
Pathology foundation models are trained on microscopic images of tissue biopsies. Such models can help us analyze many diseases, especially cancer. This enables better diagnosis, personalized treatment, and much more.
We previously released our own pathology FM, OpenMidnight!
Links with more details:
leaderboard: https://labless.dev/nano-projects/nanopath blog post: https://sophont.med/blog/nanopath code: https://github.com/MedARC-AI/nanopath
What's been interesting to see is that for the past ~3 years most pathology foundation models have converged to almost the same DINOv2 framework (with some minor modifications here or there)
There is clearly room for innovation!! This is why we have created nanopath!
The challenge is to increase this score as much as possible within a fixed training budget. To do so, you can change anything in the training recipe: arch, algorithm, regularization, optimizer, data curation, etc.
If you have a new best run, submit it to our platform!
Nanopath is a minimal training+eval framework for pathology foundation models.
The base codebase trains a DINOv2-small on the public TCGA dataset for 1 hour on a single H100.
The final score is the average of 12 diverse benchmarks that take a total of 15 min to run.
We hope nanopath can serve as an accessible test-bench to explore novel self-supervised approaches and fosters innovation in the space!
@iScienceLuvr @MedARC_AI @alokbishoyi97