Sep 15, 2026 · 43 min · 12 segments
In this episode of the American Shoulder and Elbow Surgeons Podcast, hosts Dr. Brian Waterman and Dr. Peter Chalmers interview Dr. Howard Routman and Dr. Marius Mayerhoefer about radiomics. This…
Marius MayerhoeferGuest
Howard RoutmanGuestPeter ChalmersHostBrian WatermanHostCan I, can I ask one question here that I'm just, um, trying to make sure I understand? I mean, it sounds, it, it, it sounds like, number one, you need a large data set, and you need to have some clinical outcome, and you need to have, you know, like the radiology, the, the, the actual images that you then can auto-segment and extract features from.
And then it sounds like you would be faced with, you know, a wide variety of features.
... like in the one you mentioned, entropy, I'm, I'm, I'm having a little bit of difficulty, like wrapping my mind around that.
Like, when I think of entropy, I think of like, you know, like w- what happens after my children approach a room.
... help, help me understand what does entropy mean when we're talking about a radiology, radiologic image, and are the features like interconnected or is each of them in an independent slider bar of a different way of describing an image that then...
And like what, why would, for instance, let's talk about you can use entropy or you can use a different.
Give us like a specific example of maybe in your own work, a specific radiomic feature that, that maybe surprised you was associated with an outcome that you maybe had not expected because you could take this kind of blank slate, u- u- unbiased approach.
There are, because there are so many features, um, and some of them are highly correlated with each other, especially the ones that are m- sort of members of the same family.
As I mentioned before, um, there is, for example, the category of so-called co-occurrence matrix features.
It actually just means you go looking for pixel pairs with the same gray level values.
So you have, for example, you're looking for a pixel that's black and one that is white, and they are, um, next to each other, directly next to each other in the horizontal plane, and you go looking for that exact combination, that pixel pair combination across an image, right? So, um, based on that, you can calculate different features, and one of them is entropy.
I totally understand what you mentioned with the, because I have a toddler, so I can relate, I can relate to your comment.
There is an underlying equation, a well-defined equation, how to calculate entropy.
You don't, you, you can theoretically highlight that feature across the image using a color map, so you can visualize the number across the images and show regions where the number is high and regions where this number is low.
But, um, it's just numbers that we are dealing with, and we need to try and reduce the nu- the, the number of features as much as possible by basically deselecting, removing those that are highly correlated and just leaving ideally a couple of features that are complementary and are related to your outcome.
Can I, can I ask one question here that I'm just, um, trying to make sure I understand? I mean, it sounds, it, it, it sounds like, number one, you need a large data set, and you need to have some clinical outcome, and you need to have, you know, like the radiology, the, the, the actual images that you then can auto-segment and extract features from.
And then it sounds like you would be faced with, you know, a wide variety of features.
... like in the one you mentioned, entropy, I'm, I'm, I'm having a little bit of difficulty, like wrapping my mind around that.
Like, when I think of entropy, I think of like, you know, like w- what happens after my children approach a room.
... help, help me understand what does entropy mean when we're talking about a radiology, radiologic image, and are the features like interconnected or is each of them in an independent slider bar of a different way of describing an image that then...
And like what, why would, for instance, let's talk about you can use entropy or you can use a different.
Give us like a specific example of maybe in your own work, a specific radiomic feature that, that maybe surprised you was associated with an outcome that you maybe had not expected because you could take this kind of blank slate, u- u- unbiased approach.
There are, because there are so many features, um, and some of them are highly correlated with each other, especially the ones that are m- sort of members of the same family.
As I mentioned before, um, there is, for example, the category of so-called co-occurrence matrix features.
It actually just means you go looking for pixel pairs with the same gray level values.
So you have, for example, you're looking for a pixel that's black and one that is white, and they are, um, next to each other, directly next to each other in the horizontal plane, and you go looking for that exact combination, that pixel pair combination across an image, right? So, um, based on that, you can calculate different features, and one of them is entropy.
I totally understand what you mentioned with the, because I have a toddler, so I can relate, I can relate to your comment.
There is an underlying equation, a well-defined equation, how to calculate entropy.
You don't, you, you can theoretically highlight that feature across the image using a color map, so you can visualize the number across the images and show regions where the number is high and regions where this number is low.
But, um, it's just numbers that we are dealing with, and we need to try and reduce the nu- the, the number of features as much as possible by basically deselecting, removing those that are highly correlated and just leaving ideally a couple of features that are complementary and are related to your outcome.
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.
