0:00

Forget frequently asked questions. Common sense.

0:02

Common knowledge. Or Google. How about

0:04

advice from a real genius? 95 %

0:06

of people in any profession are

0:08

good enough to be qualified and

0:10

licensed. 5 % go above and

0:12

beyond. They become very good at

0:14

what they do. But only 0 .1

0:16

% are real Jesus. Richard Jacobs

0:18

has made it his life's mission

0:20

to find them for you. He

0:22

hunts down and interviews geniuses in

0:24

every field. Sleep science, cancer, stem

0:26

cells, ketogenic diets, and more. Here

0:29

come the geniuses. is the

0:31

Finding Genius podcast. Hello,

0:38

this is Richard Jacobs with the Finding

0:40

Genius podcast, now part of the

0:42

Finding Genius Foundation. My guest today is

0:44

Maria Gritscher. We're to talk about

0:46

machine vision AI application. Maria is a

0:48

seasoned executive with over 18 years

0:50

of experience in AI driven technology and

0:52

applications. So welcome, Maria. Thanks for

0:55

coming. Thank you. Pleasure to be here. Yeah.

0:57

Tell me a bit about. your background

0:59

with machine vision AI, like, you know, what

1:01

are some really interesting things you've seen

1:03

develop over the past, you know, 18 years?

1:05

Okay, well, of course. So let me start

1:07

a little bit with my background. So

1:09

I've been, I've been working against startups all

1:11

my life. That's the only thing I've been

1:13

doing so far, knowledge of startups. So really,

1:16

it has been very interesting to see

1:18

the development of the technology and like what

1:20

we were busy with, like 15 years ago,

1:22

10 years ago, five years ago, and what

1:24

we're busy with right now. There's sort

1:26

of so much at that. The

1:34

more we live, the faster things change.

1:36

That has been true so far, at least

1:38

from what I have seen. Please

1:41

keep going with your background. What are some of

1:43

the really, I guess, interesting things you worked on,

1:45

and then we'll get to what you're working on

1:47

today. So there are multiple medical

1:49

advancements that are, well, medical

1:52

health has been an important

1:54

field in our lives so

1:56

far. And there are many,

1:58

many different verticals in medical AI

2:00

specifically that are being developed, like

2:02

medical AI or medical technology. My

2:04

favorite, I would say, my favorite

2:07

applications and also that what I

2:09

specialize in is visual AI application

2:11

and what is being developed right

2:13

now in the medical domain is

2:15

absolutely amazing and I would like

2:17

to share it in a little

2:19

bit more detail. But before we

2:21

dive into this, I just want

2:23

to do like a little comparison

2:26

to how things have been done

2:28

so far in our modern world

2:30

and how they are starting to

2:32

be done and what's actually the

2:34

difference that AI, medical AI brings,

2:36

like what makes it so unique,

2:38

what makes it so important and

2:40

incredible. Let's step in. What are

2:43

you working on right now? Okay.

2:45

So well, right now, multiple medical

2:47

AI applications. For example, for

2:49

example, if you look at the

2:51

CT scan. city scans being taken

2:53

to the hospital or x -ray

2:55

or MRI scans like everything that's

2:57

visual so those visual images so

2:59

far needed an expert, an ideologist,

3:01

cardiologist to whoever the specialist is.

3:03

They needed an expert to look

3:05

at them and to understand what's

3:07

going on and basically determine that

3:09

pathology, determine like what's, if there's

3:11

something wrong or if everything is

3:13

okay on this image. Right now,

3:15

all this is being replaced by

3:17

AI. And basically, that's kind of

3:19

looking into what we do. Keymaker,

3:21

we are a building block in

3:23

creating those types of applications. We

3:25

provide training data. We train those

3:27

models, we create custom training data

3:29

for those models, but what I

3:31

would like to elaborate is that

3:33

kind of talk a little bit

3:36

more about is how this actually,

3:38

like how different it is from

3:40

what we knew and we've been

3:42

doing so far. So what's going

3:44

on right now. So for example,

3:46

if we take, if we take

3:48

traditional medicine, how it's been working

3:50

so far. So if we have

3:52

a doctor, for example, doesn't matter,

3:54

like any specialist and the specialist,

3:56

even if it's like the best

3:58

specialist in the world, it's one

4:00

person. has a certain

4:02

number of patients he's seen so

4:04

far. He has a certain

4:06

number of years of education, things

4:08

he experienced in his life.

4:10

So it's good. So let's say

4:12

when we go and see

4:14

the specialist, we get pretty good

4:16

expertise of one person. Now,

4:18

if the same diagnosis, same expertise

4:20

is given by AI, let's

4:22

say on the same can. for

4:24

the same image here, instead

4:27

of one person, even the best

4:29

person in the world, we

4:31

tap into the intelligence and knowledge

4:33

and expertise of the hundreds

4:35

of thousands of different experts all

4:37

put together. So it takes

4:39

the whole medical examinations and understanding

4:41

of the condition to a

4:43

completely different level. So you're not

4:45

just working with the best

4:47

expert in the world. data,

4:52

let's say to identify lung cancer.

4:54

Noyans? So yes, yeah, I understand

4:56

what you mean. So it really

4:58

depends on the model. Usually there

5:00

are different models that try to

5:02

do the same different companies that

5:04

develop those models. So it takes

5:06

like thousands of images that are

5:08

annotated, multiple specialists to train a

5:10

model properly. Now wife and this

5:12

is not a process. Let's just

5:14

do it one time and that's

5:16

it. No, it's ongoing process. It's

5:18

constantly developing, which means that like

5:20

doctor has to go to courses

5:22

and go to seminars and read

5:24

papers to constantly keep his knowledge

5:26

up to date. Same with AI.

5:29

It has to keep learning. I'm

5:31

sure we're learning not from one

5:33

person, we're learning from thousands of

5:35

experts. So what kind of things

5:37

are you saying? What are some

5:39

examples? So for example, brain tumors,

5:41

identification of brain tumors early. of

5:43

brain tumors that is done immediately,

5:45

done with very, very high precision.

5:47

Another, actually, would say one of

5:49

my favorite examples is ultrasound and

5:51

ultrasound in emergency rooms. And what's

5:53

happening here, the response. The

5:55

velocity of the response is very

5:57

important, so identify if there's reflux

5:59

or any injury on ultrasound and identifying

6:02

it fast, super important to save

6:04

lives. So when we have a person

6:06

doing it, there's a delay in

6:08

the response. It also depends on the

6:10

availability of the person. But when

6:12

we have AI doing it, right

6:14

now it's starting to be implemented in

6:17

multiple emergency rooms. It's immediate and

6:19

it's way, way more precise. So here

6:21

we're actually seeing a significant impact.

6:23

measurable impact on AI systems. So I

6:25

mean, what is AI looking for?

6:27

So what is its diagnosis rate compared

6:30

to doctor? So when you're looking

6:32

at an ultrasound, it's that ultrasound, for

6:34

example, it's like a lot of

6:36

things going on. So the AI

6:38

would recognize the same reflux, the same

6:40

internal bleeding as part, let's say

6:42

part of an ultrasound, it would know

6:45

the area right away. So it

6:47

would know to recognize that this

6:49

is internal bleeding with way higher precision

6:51

accuracy and way faster than person.

6:53

Well, what is Lehi or Leitha Esther?

6:55

What is the number? Well, I

6:57

cannot tell you exactly the numbers, how

7:00

like AI, it's immediate, it's it

7:02

looks at ultrasound and it's there. Well,

7:04

for the person, it may take

7:06

a few minutes till they see it

7:08

there. And also, they might miss

7:10

it because there's always Well, what's

7:12

the background efficacy rate? What is the

7:15

rate with the AI of detection?

7:17

Like the false positives or false negatives?

7:19

I cannot tell you that exactly, but

7:22

because we develop the system, we

7:24

don't implement it, but it's significant enough

7:26

that it's proven that AI works

7:28

better. Okay. So why all of a

7:30

sudden, over the past couple of

7:32

years, does AI seem to be a

7:34

lot more advanced? What's happened in

7:36

the field where now you get things

7:39

like catGPT and reasoning modules and

7:41

all that? Why is AI... So

7:43

chat, GPT, I would say that's

7:45

not part of my domain, so I

7:47

will not be able to answer

7:49

that properly, but in terms of visual

7:52

AI, like with... With every technology,

7:54

we hit a point, in a way,

7:56

like a tipping point where we

7:58

know how to train models. We have

8:00

enough data to train models, and

8:02

it just works faster. It just takes

8:04

more training data, better performing models,

8:06

and it just works. That's technology. Are

8:08

you seeing that in your machine

8:10

learning, or what are you seeing? Yeah,

8:12

well, projects that we are working

8:14

with is machine vision AI only, so

8:16

it's only visual projects. We see

8:18

it a lot with visual projects. like

8:20

before, well, also when you develop

8:23

a model, regardless what the model is,

8:25

you develop it once and then

8:27

you retrain it. So the older the

8:29

model is, the more training models

8:31

went through, the better it gets. So

8:33

we already have like a few

8:35

years of training models and creating models.

8:37

So the models just get smarter.

8:39

It's like a brain, like human brain

8:41

that learns more and more with

8:43

time and doesn't forget. It does not

8:45

forget what happens before. It just

8:47

gets better. on,

8:52

can you tell if there is

8:54

such a thing? What do you mean?

8:56

A drift in the data. I

8:58

don't know, maybe for some reason, you

9:01

know, internal bleeding around the stomach

9:03

now is showing up differently than it

9:05

did years ago. Maybe because people's

9:07

health have changed or they're average. What

9:09

if there's a drift in the

9:11

data? So this is something that's very,

9:13

very interesting question. This is we

9:15

call it bias in the data. It

9:18

happens. There's always a bias. And

9:20

that's why we always need human input

9:22

and we always need to retrain

9:24

those models and validate that they're working

9:26

right. So let's say why we

9:28

still have human input and it's not

9:30

just 100 % automated and we stopped

9:33

training right now. How do you

9:35

identify bias? What's an example of some

9:37

bias that you've seen in your

9:39

models? We asked to figure out where

9:41

it came from. For example, let's

9:43

take a simple example. Let's look at

9:45

blood cancer. You have the cancerous

9:48

cells and sometimes regular cells. The model

9:50

can recognize regular cells as cancerous

9:52

cells by mistake. This can happen. Here

9:54

it's very important that we still

9:56

have expert input and we do validate

9:58

the model and make sure that

10:00

those false positives are marked. as such,

10:02

and then we retrain the model

10:05

on eliminating those false positives. Please

10:29

visit finiegeniuspodcast .com and click on

10:31

support us. We have three levels

10:33

of membership from $10 to $49

10:35

a month, including perks such as

10:37

the ability to see ahead in

10:40

our interview calendar and ask questions

10:42

of upcoming guests, transcripts of

10:44

podcasts you're interested in, the ability

10:46

to request specific topics or guests,

10:48

and more. Visit finiegeniuspodcast .com

10:50

and click support us today. Now

10:52

back to the show. Okay, I

10:54

mean, how many of the scans

10:56

do you look at manually to

10:58

see if it's a false positive?

11:00

Is it every 201 or is

11:02

it only? No, no, no, no,

11:04

it's usually there is ongoing process

11:06

of validation going on. So the

11:08

experts always look the data science

11:10

team usually always looks at the

11:12

outputs of the models or certain

11:14

subsets of outputs of the models

11:16

and validates them if they're correct. What

11:20

percentage of scans, for instance, for internal

11:22

bleeding or looked at manually, statistically what will

11:24

be enough? 1%, 10 %? So it would

11:26

depend on the model and it would

11:28

depend on specific use case, but it can

11:30

be anything from 10 % to if the

11:33

model is performing well, it can be

11:35

less than 1%. Okay. So again, you've gotten

11:37

it to a very high efficacy. Do

11:39

you see it getting better and better still

11:41

or? It's always going to get better

11:43

and better. It is. It's always going to

11:45

get better. I think I believe so.

11:48

I believe so, at least so far, because...

11:50

Think of it, there are always edge

11:52

cases, there are always new things that

11:54

are coming up. There's also

11:56

overtraining. I mean, there's getting stuck in

11:58

a localized minimum or maximum. you have

12:00

a speaker, it tends to dominate what's seen

12:02

in the visual field there. I mean,

12:04

again, I know there's overtraining. So how do

12:06

you make sure the model doesn't wander

12:08

into that territory? How do you tear

12:10

the weight or how do you zero

12:13

it out or make sure it's resetted? So

12:15

when the models are trained, We

12:18

start training a model, we feed

12:20

it a certain number of scans. After

12:22

some time, we need less and

12:24

less training, but we do need training

12:26

for each case or new things

12:28

that come up. So I would say

12:30

never. What

12:33

do you do? So that's where we

12:35

need a human input and then we

12:37

teach the model as we would like

12:39

as an expert in this field will

12:41

do. They would learn what it is

12:43

and explore. Basically create the training data

12:45

that outlines this edge case to the

12:48

model. What's the example of it? For

12:50

example, new types of tumors, for example,

12:52

or if you're going to the same

12:54

ultrasound where the, let's say, the organs

12:56

we're looking at look different, completely different,

12:58

or it's a person with results that

13:00

haven't been seen before. In any type

13:02

of data, H -cases happen, and that's

13:04

the thing about H -cases. We can't predict

13:06

them, they just happen, and when they

13:08

happen, then we deal with them. You

13:10

don't just delete them, I would think

13:12

you would learn from them, but do

13:14

you not include them in the training

13:16

data, or what do you do with

13:18

them? Now we have to include

13:21

them. That's how. Yeah. So in this

13:23

case, we identify it as an edge case

13:25

or the model fails. Like the model

13:27

basically says, okay, I don't know what to

13:29

do with this. This looks different. Then

13:31

we have experts, let's say radiologists or a

13:33

number of experts that look at those

13:35

edge cases and they decide what it is

13:37

and market as what they decided that

13:39

it is. And we use that as a

13:41

training data for the model. Okay. Our

13:43

edge case is particularly useful for the interest.

13:46

are extremely useful, they're essential to

13:48

keep the models up to date.

13:50

Why? How are they useful? Well,

13:52

if you encounter something new, you

13:55

want the models to learn that

13:57

something new exists. And if they

13:59

encounter this again, they will know

14:01

what to do with it. So

14:03

what kind of education would you get

14:06

with internal bleeding, for instance? Well, a

14:08

kind of hard to answer this question

14:10

that kid just doesn't look like. It's

14:13

really a specific, but let's say it

14:15

doesn't look like it's a bleeding, but

14:17

it is or vice versa, like false

14:19

negative or false positive. It can go

14:21

both ways. Okay. So that means you

14:23

edge cases tend to support. false positives

14:25

or false negatives or random.

14:28

It really depends. There is no

14:30

rule that can be anything. That's the

14:32

whole idea for edge cases. You

14:34

never know what it is, but when

14:37

you encounter model fails, then we

14:39

have to train the model on those

14:41

edge cases as well. But it

14:43

can be. There's a skew in edge

14:45

cases. They're not symmetrical around the

14:47

main data. Maybe that would

14:49

tell you something. Long tail. Yes,

14:51

but again, so it's important for

14:54

model change. Okay, so you use

14:56

AIs that detect internal bleeding when

14:58

someone comes into the hospital. What

15:00

else are they being used for,

15:02

the vision sense? There are actually

15:04

so many applications. It starts from

15:06

recognizing pathology on the scans, like

15:08

any type of scans, to cameras

15:10

in hospitals or cameras. elderly

15:12

homes that ensure that the people are

15:14

well, like they recognize that if it's

15:17

an elderly home, the camera, the smart

15:19

camera, of course, there's a full privacy

15:21

to it, but the smart camera would

15:23

be able to see if the person,

15:25

say the elderly person the house, having

15:27

an issue or they're having a heart

15:29

attack or it fell down, basically motion

15:32

recognition as well, or they need help.

15:34

So instead of basically a person to

15:36

press a button, call for help, the

15:38

camera would recognize it right away and

15:40

call the ambulance, call the sport. So

15:42

it's everything. It's really helping to, it's

15:44

really helping to, if you're looking specifically

15:47

at elderly care, it's really helping to

15:49

improve their life and save lives of

15:51

elderly people because instead of... How do

15:53

you know? Is it in use or

15:55

is it still being said? It's in

15:57

use. Yeah, yeah, it's been in use.

15:59

It's been in use actually for quite

16:01

a while. It's getting better and better,

16:04

especially for emotion recognition or action recognition. It's

16:06

usually autonomous cameras that are deployed at

16:08

people's houses, multiple companies, let's say multiple...

16:11

service providers that do that, but the

16:13

idea is this camera fully autonomous that

16:15

can recognize if the person's having like

16:17

a heart attack or stroke or anything

16:19

is wrong with them. And then it

16:21

calls for help right away. Where are

16:23

these? Only in hospitals or? You have

16:25

them in hospitals as well. You have

16:27

them in private homes. This is like

16:30

a service that you have a security

16:32

camera in a home. It's a choice

16:34

of, it's like people's choice to use

16:36

the services or not. Is that people

16:38

just having in their home? Yeah, some

16:40

people do. This technology has been around

16:42

for a few years already. Of course,

16:44

it's getting better and better. But

16:47

more and more people choose to

16:49

use that in their home. It's like

16:51

having a smart camera facing outside,

16:53

detecting motion, but the same thing, but

16:55

more sophisticated, fully private inside. Why

16:57

companies reduce this? Where can people get

16:59

it, for instance? That's a

17:01

good question. So I will not

17:03

be able to answer that. But

17:06

I'm sure if you search for

17:08

like even charge PT or Google

17:10

cameras, smart cameras for elderly care,

17:12

there are multiple companies that do

17:14

that. Now, we like keymaker, we

17:16

are a service provider for creating

17:18

the training data. So we will

17:20

be on the other end of

17:23

those cameras. We will be helping

17:25

to develop and train those models,

17:27

but we don't sell them. So

17:29

like kind of going up the

17:31

development chain, we would be like

17:33

the building block of creating this

17:35

AI and then multiple companies, multiple

17:37

camera companies or healthcare companies that

17:39

would purchase this AI as a

17:42

service together with the camera if

17:44

needed. deployed under their brand. Many

17:46

other interesting examples of work that

17:48

you've done? No, they use this

17:50

AI to diagnose. Any other examples?

17:53

There are so many. Let me

17:55

think for a second. What else

17:57

is really interesting? I

17:59

would say of the examples

18:01

I personally really like is goes

18:03

back to what I mentioned before

18:05

is x -ray or MRI recognition. If

18:07

there's any issue with a broken

18:09

bone or internal bleeding or tumors,

18:11

etc. But the way it's used

18:13

is in remote locations. So you

18:15

have, let's say, an x -ray

18:18

scanner in a remote location, either

18:20

in places that are just, say,

18:22

far away. There's no hospitals there,

18:24

no doctors there, but there's the

18:26

Khalil station or the medics station

18:28

with those, with the scanning device.

18:30

And instead of having waiting for

18:32

a doctor or transporting yourself to

18:34

a hospital, the diagnosis can be

18:36

done right away at the spot.

18:38

So it truly enables fast and

18:40

cheap healthcare in a way in

18:42

remote location. Now, the most, just

18:44

so far, like the most important

18:46

part of the health, like of

18:48

helping somebody is understanding what's wrong.

18:50

And here, we immediately understand what's

18:52

wrong and choose, even remotely, choose

18:54

the appropriate treatment for that person.

18:56

So it really helps to help

18:58

people to get proper health care

19:01

in the remote or locations first.

19:03

Nothing's near there, no hospitals, no

19:05

anything. Okay. What's the best way

19:07

for people to keep tabs on

19:09

your work? Where should they go?

19:11

Well, again, like chemo care, we're

19:13

not at the end of... creator

19:15

with just a building block and

19:17

creating those models. So nothing really,

19:19

but there's also news and things,

19:21

new things come up and healthcare,

19:23

so just general stuff. Okay, well,

19:25

very good. Well, thank you for

19:27

coming on the podcast and explaining

19:29

it. I really appreciate it, Maria.

19:31

Thank you. If you like this

19:33

podcast, please click the link in the

19:35

description to subscribe and review us on

19:37

iTunes. You've

19:42

been listening to the Finding

19:44

Genius podcast with Richard Jacobs. If

19:47

you like what you hear, be

19:49

sure to review and subscribe to

19:51

the Finding Genius podcast on iTunes

19:53

or wherever you listen to podcasts.

19:55

And want to be smarter than

19:57

everybody else? Become a premium member

19:59

at .VendingGeniusPodcast .com. This podcast is for

20:01

information only. No advice of any

20:03

kind is being given. Any action

20:06

you take or don't take as

20:08

a result of listening is your

20:10

sole responsibility. Consult professionals when advice

20:12

is needed.