Founder of Datamint, an end-to-end research platform for medical AI. A man of simple needs: an AA surface plate, a fast oscilloscope, an a stock of Malbec.
Edmonton, Alberta
Joined October 2008
- Tweets725
- Following65
- Followers96
- Likes11
I'm continuing a series consolidating machine learning practices and vocabulary for non-AI specialist clinical researchers. This one talks about evaluating segmentation models.
learn.datamint.io/articles/s…
Rod Fitzsimmons Frey retweeted
Americans are more negative than positive about AI’s impact across nearly every area surveyed.
The one exception: healthcare.
We have a unique opportunity to showcase the benefits and set the standard for how AI can meaningfully improve people’s lives.
I'm writing a series of articles for clinical researchers who have been so focused on the esophagus or whatever that they somehow failed to get a parallel degree in machine learning. First up: metrics for evaluating models. learn.datamint.io/articles/c…
Obviously AI researchers want their PhDs and postdocs to succeed. But when they walk out the door for industry, the project instantly falls 6 months behind because so much project structure is implicit and lives in their brain or on their laptop. Centralize your data, people!
Of course, you don't have to worry, because they promised they'd be available by email if there were any questions. 🙃
datamint.io/post/how-central…
Clinical AI researchers want to stretch their money, just like every other human. It's getting common to do image segmentation (feeding into classification models) with foundation models. That's probably hurting their results. datamint.io/post/annotation-…
My list of things I can do better than AI is getting shorter. I'm not sure what's more concerning: that the list is shrinking or that I decided I needed a list
Trying to learn marketing in 2026 feels like taking up fine woodworking the year the Ikea opened in your town.
4/4 So in a project with limited resources (is there any other kind?), improving annotation quality is often a better spend than collecting more scans.
Longer version, with the reasoning: datamint.io/post/annotation-…
3/4 That makes the error systematic. Random error averages out as you add data, but systematic error doesn't. Your model just gets more confident about the cases the readers got wrong. It gets arrogant rather than better.
2/4 Ground truth in medical imaging is rarely a clinical outcome or biopsy result. It's usually expert judgement, and experts disagree. Not randomly, though, they disagree in a consistent way, usually on the same borderline cases.
1/4 "How many images will I need?" is the first question I get when onboarding a researcher into Datamint.
The flip answer is "lots". But that’s not completely true. They might already have enough.
I may reduce my odds of getting sick from COVID-19 by not touching my face, etc.
Downside is I’ll probably be unemployed, since stroking my beard and looking thoughtful is my singular business skill.
Turns out we have solid historical examples of nations decarbonizing. Like France in the 1980s.
You’re unable to view this Post because this account owner limits who can view their Posts. Learn more
I'm *really enjoying* writing in Clojure, but I can't shake the feeling I'm doing it just to make myself feel smart.
If you start your comment with "umm... ", "you do know that...", "of course you realize..." etc then fuck you.
Every 18 months or so I lose 3 days productivity to @richhickey videos. #clojure