Murat Seçkin Ayhan

Literally! 😃

It’s a beautiful day. 🧿🍀🤞

Turtwig would be nice. :)

One of the followings, in the order of preference, will be the name of my new GPU server:

1. “turing”: to remember Alan Turing.
2. “heidelberg”: one of the most beautiful cities in Germany.
3. “dartmouth”: for Dartmouth College - the birthplace of AI.

What is yours? 😃

I hope they all find you well. 😃

I have moved to the breathtaking Pacific Northwest and joined the ranks of the University of Washington in Bothell.

Here’s to new opportunities, growth, and discovery! 🥂🧿🍀

Thanks and congratulations to the team, especially my partner in crime David Merle 👁️👨‍⚕️

I would to love see what can be done for other diseases or use cases. With even more capable models like the Gemini 2.5 instances, the best is yet to come! 😉

In addition to textual outputs for interpretability, our prompts led to fairly well-calibrated uncertainty estimates from Gemini out of the box. Reliable uncertainty measures can help clinicians judge whether automated decisions can be trusted and integrated into their workflow.

This can be done while also providing explanations and counterfactual insights into the model’s decision in natural language space, which is useful for clinicians.

In-context learning for data-efficient diabetic retinopathy detection via multimodal foundation models This study aims to evaluate whether in-context learning, a prompt-based learning mechanism enabling multimodal foundation models to rapidly adapt to n…

🚨 New paper alert 🚨

Can we match RETFound’s retinal disease detection performance by just prompting a general-purpose foundation model like Gemini 1.5 Pro to the task? For diabetic retinopathy detection from color fundus photos, YES!

www.sciencedirect.com/science/arti...

Thanks and congratulations to the team, especially my partner in crime David Merle 👁️👨‍⚕️

I would to love see what can be done for other diseases or use cases. With even more capable models like the Gemini 2.5 instances, the best is yet to come! 😉

In addition to textual outputs for interpretability, our prompts led to fairly well-calibrated uncertainty estimates from Gemini out of the box. Reliable uncertainty measures can help clinicians judge whether automated decisions can be trusted and integrated into their workflow.

This can be done while also providing explanations and counterfactual insights into the model’s decision in natural language space, which is useful for clinicians.

I need GPUs, lots of GPUs. 🔥🤓

Realistic retinal image generation for counterfactual reasoning in ophthalmology. Diffusion models coupled with robust classifiers led to stunning results. It works on both color fundus photographs and OCT.

Finally out in PLOS Digital Health!

In-context learning for data-efficient classification of diabetic retinopathy with multimodal foundation models Importance: In-context learning, a prompt-based learning mechanism that enables multimodal foundation models to adapt to new tasks, can eliminate the need for retraining or large annotated datasets. W...

We matched the performance of RETFound for DR detection from CFPs via in-context learning with Gemini 1.5 Pro. We also achieved counterfactual reasoning about diagnostic decisions in natural language space, plus well-calibrated predictive probabilities.

There goes my very first like on BlueSky!