Sep 28, 2026 · 48 min · 11 segments
Send us Fan Mail We are all affected by health care, including our own health and the health of friends and family. Did you know that there are…
Jon GiambattistaGuest
David de MontignyHost
We didn't start with the intention of like building a company or anything.
We just thought, you know, let's try to solve this problem and see what we can do.
Because all the while that you're sort of, I don't want to say tinkering around with it, but like trying to learn this new technology, you're still working at Sask Energy.
You've got your full-time job and you're doing that, but this is kind of like in the evenings and on the weekends and seeing if we can actually make something that works.
And yeah, at the start, we really had no idea where to even start or what we were doing.
And so Josh found a bunch of open source public domain CT images from, I think, various clinical trials.
And we started off just kind of as a proof of concept to see if we could train a model for bladder.
Like, can we build a model that would draw contours around bladder reasonably well? And so Josh put together this initial training data set of hand-drawn contours.
And pretty immediately we realized that like this technology was going to work for this application.
Like it was obvious in the sense that we, we started seeing, we started trying the model that we trained on, you know, patient or patient data that wasn't in the training data set.
And it was really amazing how well it was generalizing to the variability and anatomy.
And, and, Yeah, just to kind of explain it, if you're looking at a CT scan and you're trying to like segment or contour a bladder, every patient is going to have different anatomy.
Like they'll have different amounts of like air in their bowels, their bladders might be empty or full.
If they have cancer, they might have a tumor that's like adjacent to the bladder and distorting that anatomy or maybe they'll have a tumor that's far away.
We didn't have that much data in this training data set at all, but we saw that our model was adapting to all of this new variability and still performing very well.
We didn't start with the intention of like building a company or anything.
We just thought, you know, let's try to solve this problem and see what we can do.
Because all the while that you're sort of, I don't want to say tinkering around with it, but like trying to learn this new technology, you're still working at Sask Energy.
You've got your full-time job and you're doing that, but this is kind of like in the evenings and on the weekends and seeing if we can actually make something that works.
And yeah, at the start, we really had no idea where to even start or what we were doing.
And so Josh found a bunch of open source public domain CT images from, I think, various clinical trials.
And we started off just kind of as a proof of concept to see if we could train a model for bladder.
Like, can we build a model that would draw contours around bladder reasonably well? And so Josh put together this initial training data set of hand-drawn contours.
And pretty immediately we realized that like this technology was going to work for this application.
Like it was obvious in the sense that we, we started seeing, we started trying the model that we trained on, you know, patient or patient data that wasn't in the training data set.
And it was really amazing how well it was generalizing to the variability and anatomy.
And, and, Yeah, just to kind of explain it, if you're looking at a CT scan and you're trying to like segment or contour a bladder, every patient is going to have different anatomy.
Like they'll have different amounts of like air in their bowels, their bladders might be empty or full.
If they have cancer, they might have a tumor that's like adjacent to the bladder and distorting that anatomy or maybe they'll have a tumor that's far away.
We didn't have that much data in this training data set at all, but we saw that our model was adapting to all of this new variability and still performing very well.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.
