rohan

🧵 Everyone is chasing new diffusion models—but what about the representations they model from?
We introduce Discrete Latent Codes (DLCs):
- Discrete representation for diffusion models
- Uncond. gen. SOTA FID (1.59 on ImageNet)
- Compositional generation
- Integrates with LLM
🧱

Presenting ✨ 𝐂𝐇𝐀𝐒𝐄: 𝐆𝐞𝐧𝐞𝐫𝐚𝐭𝐢𝐧𝐠 𝐜𝐡𝐚𝐥𝐥𝐞𝐧𝐠𝐢𝐧𝐠 𝐬𝐲𝐧𝐭𝐡𝐞𝐭𝐢𝐜 𝐝𝐚𝐭𝐚 𝐟𝐨𝐫 𝐞𝐯𝐚𝐥𝐮𝐚𝐭𝐢𝐨𝐧 ✨

Work w/ fantastic advisors Dima Bahdanau and @sivareddyg.bsky.social

Thread 🧵:

WSJ sub: “Indian Partition went smoothly enough, right? Will try to remember to check this before publishing.”

8/ The method is open-source and available as part of the Spinal Cord Toolbox (SCT)🚀
Have questions or feedback? Let us know! We’d love to hear how EPISeg can support your research.

7/ This work wouldn’t have been possible without the amazing contributions from multi-site collaborators, open-source community, @PolyNeuro , @mila-quebec.bsky.social and @StanfordPain . We hope EPISeg helps research paving the way for new insights into spinal cord function and dysfunction.

6/ What sets EPISeg apart?
• Trained via active learning, improving robustness over multiple iterations.
• Works across different scanner protocols and resolutions.
• Handles clinical cases, including data from patients with myelopathy and fibromyalgia.

5/ To build EPISeg, we created an open-access dataset of SC fMRI along with the segmentations from 15 sites, covering 406 participants with varying scanner setups and conditions.

4/ Our solution: EPISeg uses a deep learning model trained on diverse, multi-center dataset of gradient-echo EPI images to perform fully automated SC segmentation. It’s fast, accurate, and robust to:
• Low-resolution images
• Distortions
• Signal drop-outs
• Motion artifacts

3/ Until now, spinal cord (SC) segmentation on EPI data required time-consuming manual corrections prone to user bias and errors.

2/ Spinal cord fMRI is critical for studying sensation, movement, and autonomic function. However, preprocessing SC fMRI data like segmenting the spinal cord is challenging due to low spatial resolution, susceptibility artifacts, motion and ghosting artifacts and poor SC contrast.

Workflow of EPISeg

📢New preprint and data alert!
1/ EPISeg: Automated segmentation of the spinal cord on echo planar images using open-access multi-center data
Preprint: www.biorxiv.org/content/10.1...
Dataset: openneuro.org/datasets/ds0...

Now accepted at @ismrm.bsky.social 25 for power pitch oral
More in thread 🧵

6/ What sets EPISeg apart?
• Trained via active learning, improving robustness over multiple iterations.
• Works across different scanner protocols and resolutions.
• Handles clinical cases, including data from patients with myelopathy and fibromyalgia.

5/ To build EPISeg, we created an open-access dataset of SC fMRI along with the segmentations from 15 sites, covering 406 participants with varying scanner setups and conditions.

4/ Our solution: EPISeg uses a deep learning model trained on diverse, multi-center dataset of gradient-echo EPI images to perform fully automated SC segmentation. It’s fast, accurate, and robust to:
• Low-resolution images
• Distortions
• Signal drop-outs
• Motion artifacts

3/ Until now, spinal cord (SC) segmentation on EPI data required time-consuming manual corrections prone to user bias and errors.

2/ Spinal cord fMRI is critical for studying sensation, movement, and autonomic function. However, preprocessing SC fMRI data like segmenting the spinal cord is challenging due to low spatial resolution, susceptibility artifacts, motion and ghosting artifacts and poor SC contrast

has anyone found an efficient way to bookmark posts here?

Coming soon to Gradio Chatbot!

happy that my feed still has cat posts

I'm making an unofficial starter pack with some of my colleagues at Mila. WIP for now but here's the link!

go.bsky.app/BHKxoss

are you looking for specific areas in cv + medical ai? or just latest papers?