0:02

Hello and welcome to noon conferences hosted

0:04

by MRI online. Noon conference was created when

0:07

the pandemic hit as a way to connect the global

0:09

radiology community through free live educational

0:12

conferences that are accessible for all.

0:15

It has become an amazing weekly opportunity to

0:17

learn alongside radiologists from around the

0:19

world and we encourage you to ask questions and

0:21

share ideas to help the community learn and grow.

0:24

You can access the recording of today's

0:26

conference and previous noon conferences

0:28

by creating a free MRI online account.

0:30

The link will be provided in the chat box.

0:32

You can also sign up for a free trial of

0:34

MRI Online Premium Membership to get access

0:37

to hundreds of case based learning courses

0:40

across all key radiologies and specialties.

0:43

Learn more at mrinline.

0:45

com.

0:46

Today we're honored to welcome Dr.

0:47

Rachel Brehm, a leader in the field of breast imaging.

0:51

For a lecture on AI and mammography

0:53

leveling healthcare disparity.

0:55

Dr.

0:55

Brehm is a fellow of the American College of

0:57

Radiology and the Society of Breast Imaging.

1:00

She is currently the Vice Chair of Radiology

1:03

and the Director of Breast Imaging and

1:05

Intervention at George Washington University

1:07

School of Medicine and Health Sciences.

1:10

Since joining the faculty of GW, Dr.

1:13

Brim and her colleagues have developed a state

1:15

of the art patient focused breast imaging center

1:17

that integrates compassionate care with the latest

1:20

technology, including full digital technology and

1:23

CAD, as well as MR and molecular breast imaging.

1:28

Dr.

1:28

Brim's commitment to changing the face of breast

1:30

cancer is both a professional and personal passion.

1:33

She is the director of GW Mobile Mammography

1:35

Program, which focuses on screening

1:37

mammography in the underserved community in D.

1:40

C., Maryland, and Virginia.

1:43

She is the chief medical officer for the

1:44

non profit Brim Foundation, and she works

1:46

extensively with members of Capitol Hill to

1:49

further education, awareness, and legislation to

1:51

optimize breast cancer diagnosis and treatment.

1:54

As a breast cancer survivor herself, she

1:56

understands and is committed to the science

1:58

and the personal aspects of breast cancer.

2:01

We look forward to hearing from her today

2:03

about the ways that technology can help

2:04

close global disparity gaps in breast care.

2:08

At the end of the lecture, join Dr.

2:10

Brim in a Q& A session where she will address

2:12

questions you may have on today's topic.

2:14

Please use the Q& A chat feature to

2:16

submit your questions and we will get to

2:17

as many as we can before our time is up.

2:19

With that being said, we welcome you.

2:21

Dr.

2:21

Brenn, please take it from here.

2:23

Thank you very much.

2:25

I am absolutely delighted to be here today

2:27

and to share with you some of the, uh, most

2:30

exciting tech, uh, changes, uh, in mammography

2:36

and in approaches to leveling the playing field

2:39

of unacceptable, uh, Healthcare disparities,

2:42

both in the United States and elsewhere.

2:45

So let's start out by talking about

2:47

what artificial intelligence is.

2:49

And really, it's an umbrella term that we use

2:52

to describe using machine learning algorithms

2:55

and other means of technology to make decisions.

2:59

So we end up with With cognitive decisions.

3:02

So really, in the simplest sense, it's when computers

3:04

and machines mimic how we think and how we decide on,

3:10

uh, decisions in terms of health care or anything else.

3:13

But it's decisions are taking actions

3:16

based on the output of a computer.

3:20

And if you look at some of the recent,

3:22

uh, Publications by Frost and Sullivan.

3:25

They project that artificial intelligence

3:27

in cognitive computing are going to really

3:31

change health care and change medicine by

3:34

2025 with an industry over 150 billion.

3:39

And it's also estimated that if AI is

3:42

implemented correctly, it could improve

3:44

the health care outcomes by 40 percent and

3:47

reduce treatment treatment costs by half.

3:50

by improving the diagnosis, increasing

3:52

access and enabling precision medicine,

3:55

which is where we're going today.

3:58

So why are we talking about AI in terms of mammography?

4:01

Mammography is a pretty good examination.

4:03

Over the past two decades, we've seen, we've seen an

4:06

extraordinary reduction in the mortality from breast

4:09

cancer with an over 40 percent decrease in the death

4:12

rate from breast cancer over the past two decades.

4:15

And that is significantly impacted

4:18

by the widespread use of mammography.

4:21

So that's the good news.

4:22

But we also know that mammography

4:24

is an imperfect examination.

4:26

The overall sensitivity of mammography

4:28

for breast cancer detection is 85%.

4:31

But up to up to 50 percent of breast

4:33

cancers are, I wouldn't say missed.

4:36

I'd say mass.

4:37

by dense breast tissue on a mammogram.

4:41

And we've also seen that although we have

4:43

had great strides in mortality reduction,

4:47

it's not equal across all populations.

4:50

And that the mortality reduction in black American

4:53

women is 40 percent less than white American women.

4:57

And so we also know that access to

4:59

mammography is not equal for everyone.

5:02

So what we can do is we can use AI.

5:07

Sorry, uh, to improve the sensitivity and specificity

5:12

in less time so we can be better, faster, quicker,

5:17

and we can harness the enormous genomic information

5:20

that we know exists in mammography and other

5:22

imaging modalities MRI and even ultrasound for

5:26

more accurate individualized risk assessment.

5:29

And perhaps most in the near future, we can

5:33

use AI to increase the availability of not just

5:37

mammography, but the highest quality mammography

5:40

and interpretation, both in the US and globally.

5:44

And with that.

5:45

dramatically improve the outcome from breast cancer.

5:50

So here's a map of the United States

5:52

and you can, I'm sorry, of the world.

5:54

And this is at the top cancers

5:56

in women across the world.

5:58

And you can see that pink, that breast cancer

6:01

overwhelmingly is the most common cancer in virtually

6:04

every country globally, except for a few countries

6:07

in Africa, which I think are really important.

6:10

We'll talk about that in a second.

6:11

Thank you.

6:11

One country in South America and non

6:14

melanoma skin cancer in Australia.

6:17

And here we see the incidence

6:19

of breast cancer globally.

6:21

You can see that in North America, Canada, and the U.

6:24

S.

6:24

there's very high incidence of breast cancer.

6:26

Australia, Western Europe, across the globe,

6:29

we see pockets of very high incidence.

6:33

But I think what's really stark and really

6:36

daunting is look at the mortality, the

6:38

incidence of breast cancer death rates.

6:41

And you can see that in some parts of the world,

6:44

like North America, where the incidence is

6:47

very high, the death rate does not follow suit.

6:50

And similarly, or conversely, um, in Africa, where the

6:55

incidence is quite low, the death rate is very high.

6:59

So if you look at the incidence and the death rate.

7:02

They if resources were equal, they

7:05

should parallel, but they don't.

7:07

And this is a really stark reality of the

7:10

global disparity in breast cancer that we see.

7:16

So, what do we need, well the most.

7:19

Dear resource is human resource.

7:22

We really do have the global opportunity to set

7:25

up machines even in rural parts of the world.

7:28

But we don't have the people to interpret

7:31

the mammograms, the life saving screening

7:33

mammograms that we could offer in parts

7:36

of the world that don't currently have it.

7:38

So how do we leverage a I to

7:40

address this really critical need?

7:43

And then how do we move beyond that and say not only

7:45

how do we increase the availability of screening

7:48

mammography, but how do we optimize radiologist

7:52

interpretation or any interpretation of mammography,

7:56

because as we'll show scientifically and as we know,

7:59

not every person interprets mammograms equally.

8:02

So how do we answer the critical radiologist

8:05

breast image or shortage globally.

8:08

Now, lest you think this is only a

8:10

global issue, it's not the case at all.

8:12

Here's a map of the United States and you can see

8:15

that implementation of screening varies widely in

8:19

New England, it's almost 83 percent of women undergo

8:23

screening mammography in the mountain states.

8:26

It's 73%.

8:28

So we see large geographic variabilities in,

8:32

in, um, screening mammography utilization.

8:36

And then if we look at some further into some of

8:38

the healthcare disparities, um, Black women have a

8:42

nearly 40 percent higher age adjusted breast cancer

8:45

death rate than white women in the United States.

8:48

And the relative risk of death in black Americans is 71

8:51

percent higher, um, and 14 percent higher in Hispanics.

8:57

We know that black women undergo

8:58

tomosynthesis at a significantly lower

9:01

rate than white women in the United States.

9:03

And we also know that the time from diagnosis to

9:06

initiation of treatment in women with newly diagnosed

9:09

breast cancer is significantly longer for black women.

9:13

So we have a lot of challenges

9:15

here that we really have.

9:17

Moral, um, and medical obligations to impact.

9:23

So let's look at some clues as to how we can do this.

9:26

Black race is generally associated with the

9:28

lower socio economic job geographic location.

9:33

And we also know that being in

9:35

a lower socio economic location.

9:37

Reduces the access to quality health care.

9:40

Now, interestingly, this is not as true for

9:42

Hispanics in the United States, where their

9:45

prognosis is, although associated with lower

9:49

socio socio economic conditions, they have

9:53

a better prognosis than white Americans.

9:55

So unclear why that is, but let's focus on trying to

9:58

get some clues as to how to impact and do all we can

10:03

to wipe out these unacceptable health care disparities.

10:07

So something very interesting happened in Chicago

10:11

in, in, in around the year 2000, early 2000s.

10:15

What happened is they initiated a breast

10:17

cancer health disparities program.

10:19

And what they did is they said, well, what

10:22

about if we don't just offer mammography.

10:25

or, or breast cancer diagnosis and treatment.

10:29

But what about if we order, if we offer quality

10:33

care, we don't just check the box by having

10:37

mammography, but we go to ACR centers of excellence.

10:40

Or what about if we go to NCI

10:42

centers for breast cancer care?

10:44

And the reason they started looking at this in

10:46

Chicago is because they found A lot of variation and

10:50

fragmentation of care with regard to breast cancer,

10:54

particularly on Chicago South side, which they felt

10:57

affected the diagnosis and stage of breast cancer.

11:02

And so what they developed is something

11:04

called the Metropolitan Chicago breast

11:06

cancer task force, which still exists.

11:08

But it's now called equal hope.

11:10

And what happened is it's a not for profit organization

11:13

that was set up in 2008 after they found that

11:16

there was marked difference in the mortality rate.

11:20

Of breast cancer in black and white women in

11:23

Chicago, and from 2005 to 2007 black women in

11:28

Chicago died of breast cancer at a 62 percent higher

11:31

rate than one than white women, even though in

11:35

other large cities like New York and San Francisco

11:38

that had large African American populations, that

11:42

disparity was not as high as it was in Chicago.

11:46

And so they felt that this was likely due to structural

11:50

racism unavailability of care in equitable health care.

11:55

And so they developed this system where they set

11:57

up patient navigators with women and underserved

12:00

communities in the south side of Chicago,

12:02

and they went to these centers of excellence.

12:05

Not again, not just here's mammography,

12:08

but here's quality mammography.

12:10

Here's not just breast cancer care.

12:12

Here's quality breast cancer quick care.

12:15

And so they then looked at this, um, later on in 2013.

12:20

And what they found is in all the large major

12:23

cities in the United States, the disparity in breast

12:26

cancer mortality increased over this time period.

12:29

Except in Chicago, where it decreased, and

12:32

they concluded that it was not just the amount

12:35

but the quality of care to African American

12:37

women in Chicago that made the difference.

12:41

And that if you can elevate the quality of

12:43

mammographic interpretation consistently,

12:46

both in the US and globally, you can

12:48

only imagine what the impact would be.

12:51

So I think this is a place that we can really

12:54

think about using artificial intelligence to offer

12:57

not just care, but excellent quality of care.

13:00

And we talked about some of the

13:04

current healthcare disparities.

13:06

There are fewer ACR centers of excellence

13:09

of American College of Radiology centers

13:11

of excellence in underserved communities,

13:14

which are more often predominantly black.

13:16

And as we know, and as we're trying to impact

13:19

fewer black women participate in clinical trials.

13:22

which doesn't give us the opportunity to really

13:26

have more genomic and personalized medicine in these

13:29

populations that don't participate in clinical trials.

13:32

So what's the opportunity for AI in mammography?

13:37

Well, there's an opportunity and let's see if

13:40

we can improve the accuracy and decrease the

13:42

variability of interpretation, decrease the

13:45

number of false negative and false positives.

13:48

And decrease the number of interval cancers.

13:50

And as we'll say again, interval cancers

13:53

are larger, higher grade, higher stage, and

13:57

poor prognosis than non interval cancers.

14:00

Those are cancers that are diagnosed

14:02

within a year of a screening mammogram.

14:05

Can we improve the efficiency

14:08

of mammographic interpretation?

14:10

And can we triage cases by complexity?

14:13

And in Europe, where Virtually all countries that have

14:17

screening mammography programs require a second reader.

14:21

Can we invoke AI as the second reader

14:23

to make more mammography available?

14:27

Can we decrease the interpretation time and allow us

14:30

to increase the number of mammograms we interpret.

14:33

And what about the kind of, um, you know,

14:36

transition transformational idea of standalone

14:38

AI, what about mammography being interpreted

14:41

by computer without ever passing the eyes of

14:44

a radiologist so that's very complex with the

14:47

malpractice environment in the United States but.

14:50

What about the opportunity to impact mammography

14:53

in those places in Africa that we started

14:55

this talk with, um, with using AI alone?

14:59

And as we go to precision screening and

15:01

precision risk assessment, how can we harness

15:05

the enormous genomic and proteomic data available

15:08

in mammograms and other imaging modalities

15:11

to individualize screen risk assessment?

15:15

And move forward from population based risk assessment

15:18

to truly individualized genomic risk assessment.

15:22

So let's let's try to examine that.

15:25

What about when the radiologist uses artificial

15:29

intelligence in interpreting mammograms.

15:32

So here's a study that was published,

15:34

looking at 242 D digital mammograms.

15:38

interpreted by 14 radiologists with and without AI.

15:42

And, uh, the outcomes were the, uh,

15:44

area under the curve, the specificity

15:46

sensitivity and the reading time.

15:49

And here you can see the result of this study.

15:51

And there was a statistically significant increase

15:54

in performance in the area under the curve

15:57

when radiologists used AI with 2D mammograms.

16:02

And now I just want to point out that,

16:04

like so many things with computers,

16:06

AI is continuously involving evolving.

16:10

And this dashed blue line is the line of

16:13

the data, the algorithm that was used at

16:15

the time that this study was performed.

16:18

And you can see that with improvements,

16:21

there is further improved performance

16:25

of AI, it with the straight with the.

16:28

With the blue line.

16:30

So, um, as we move forward and as we talk about

16:33

data, we should also understand that this is a

16:35

constantly involving and improving, um, situation.

16:39

So if we were using the current algorithms in this

16:43

2D study, the improvement in, in, um, performance

16:48

would likely be substantially even greater.

16:52

What about with 3D mammography with tomosynthesis?

16:55

And here you can see another study looking

16:57

at 240 3D mammograms read by 18 MQSA

17:01

qualified radiologists once and then again.

17:04

And you can see in the dark blue dots representing

17:07

each radiologist, the dots moved up and to the left,

17:12

meaning that the performance improved, um, when, uh,

17:15

radiologists used AI with 3D mammography as well.

17:21

Now here's a terrific study that was published

17:23

by Emily Conant and her colleagues, and it's sort

17:26

of the Holy Grail, it shows that AI is really the

17:29

Holy Grail, because what they did is they looked

17:32

at 3d mammography with AI, and that the area

17:36

under the curve improved by significantly well 5.

17:40

7%.

17:42

That the sensitivity improved that in 21 of 24

17:46

radiologists, the specific the sensitivity improved.

17:50

But I think what's really outstanding is that the

17:52

specificity improved because generally when we

17:55

find more cancers when our sensitivity improves.

17:59

We also find more benign things and our specificity

18:02

decreases, but here that's not what happened.

18:05

What happened is the sensitivity improved, the

18:08

specificity improved, that even subspecialists,

18:12

breast radiologists improved, that the recall

18:15

rate decreased, and the reading time decreased.

18:18

So what more can you ask for?

18:20

Improvement in sensitivity, improvement

18:22

in specificity, lowering the recall rate.

18:25

Decreasing the reading time, and this was true not only

18:28

for general radiologists, for breast imagers as well.

18:30

So, pretty amazing, and here you can see that

18:33

the sensitivity improved and the specificity

18:36

did not, uh, quite achieve a P of 0.

18:39

05, but it did achieve a P of 0.

18:42

06, so it trended positively.

18:44

So being a radiologist, I couldn't possibly give

18:47

this talk without at least showing one case.

18:49

And here you can see a case of a cancer that

18:52

was, had a high score on AI, and this is how it's

18:57

marked and it was not appreciated by radiologists.

19:01

So, AI can improve the accuracy

19:03

of interpretation of mammography.

19:05

This is true both with full field mammography,

19:07

2D mammography, and 3D mammography.

19:10

And you know, really powerfully, even experienced

19:14

breast imagers have improvement in sensitivity,

19:17

but more improvement with general radiologists.

19:21

So what about the reading time?

19:24

So you can see here in one study that using

19:27

one AI that you can see that most radiologists

19:30

had a reduced reading time with AI support.

19:33

And what's interesting is that the reduction

19:36

in reading time was even greater, the

19:38

greatest with the least complex mammograms.

19:41

And that's what this curve shows.

19:43

So there's a higher reduction

19:45

for low, um, scores, AI scores.

19:49

And what that means is that you could really get

19:50

through a lot of normal mammograms very quickly.

19:54

And here, in a study by Emily Conant, again,

19:57

you can see that there is a lot of variability.

20:00

The lighter gray is the time

20:02

for interpretation without AI.

20:04

The darker gray is the time for interpretation with AI.

20:07

So for some readers, uh, Reader 1,

20:10

there was little decrease in time.

20:13

But for some readers like reader nine, there was a

20:15

marked decrease in interpretation time when AI is used.

20:21

for interpretation.

20:22

So we have significant improvement or decrease in

20:27

the time for interpretation of Tomo, uh, and it's

20:31

greatest with the least experienced radiologists.

20:35

So you get more accuracy and it's decreased

20:38

interpretation times even more so with lower complexity

20:41

cases, which allows you to get through more normal

20:44

mammograms and more availability of radiologists.

20:48

What about variability and interpretation?

20:51

Well, most mammograms, both in the United States and

20:54

globally, are interpreted by general radiologists.

20:57

And there was a fascinating article that

20:59

was recently published by the group at NYU

21:04

that said, well, what is the variability?

21:06

What are the characteristics of

21:09

performance of screening mammography?

21:11

And what they found is quite fascinating, they

21:14

found that there is geographic variability

21:17

in, uh, performance of screening mammography

21:20

interpretation, that women interpret mammograms

21:24

better than men, that academic radiologists are

21:27

better, have a better performance than non academic,

21:31

and the last two, not surprisingly, that breast

21:33

imagers who do virtually all breast imaging have

21:37

a better performance than general radiologists.

21:39

and more experienced radiologists perform

21:42

better than less experienced radiologists.

21:44

So can we use AI to level this playing field?

21:49

Can we, um, use AI to make

21:52

everybody in an acceptable range?

21:54

And this was really kind of a daunting fact.

21:56

You can see for the cancer detection rate in this

22:00

paper, um, That only 70 percent 77 percent of

22:04

radiologists were in the acceptable range of 2.

22:07

5 cancers per thousand screening mammograms.

22:11

And we know that that range should be between two

22:14

and five or six or seven cancers per thousand.

22:17

But there was a substantial percentage or

22:20

number of radiologists in the study that did not

22:23

perform, did not function in the acceptable range.

22:26

And look at the PPV2 recommendation for biopsy,

22:30

and, um, in that case, only slightly more

22:33

than 50 percent of the radiologists actually,

22:36

um, performed at, in the acceptable range.

22:40

And, uh, you know, again, there's room for

22:42

improvement and for making everybody who interprets

22:46

mammography perform in that, um, And you can see

22:50

that there was an improvement in sensitivity, both

22:52

for breast radiologists and general radiologists.

22:55

But I think it's so here you can see that

22:57

without AI, it was 81 percent to 71%.

23:01

But I think what's really so exciting is that The

23:05

look at the difference in sensitivity and the ability

23:07

to detect breast cancer in breast subspecialists

23:10

and general radiologists with the use of AI.

23:13

It's virtually the same.

23:15

So, AI can elevate the performance

23:19

of general radiologists.

23:21

sensitivity for cancer detection to

23:23

that of breast radiologists with AI.

23:27

So I think that's really exciting.

23:30

And here again, in another paper that looked

23:32

at impact of AI interpretation of breast

23:34

specialists and general radiologists, there's a

23:37

lot of lines here, but if you look, you can see

23:40

that the blue lines are general radiologists.

23:43

And the orange lines are breast radiologists.

23:47

And with, and this is AI alone is the green line above.

23:51

So you can see that without AI, the blue lines

23:55

are lower than the orange lines, meaning that the

23:58

general radiologists are performing at a lower level,

24:01

lower interpretation than, uh, than breast imagery.

24:06

But if you move over with, uh, AI, You can

24:09

see that the lines overlap that the cat that

24:12

the AI line the green line overlaps with the

24:16

performance of breast radiologists overlaps

24:19

with the performance of general radiologists.

24:22

So the answer is, we have proof here

24:25

that general radiologists can perform

24:27

like some specialized breast imager.

24:30

With the use of A.

24:31

I.

24:32

And here's another example of

24:35

the impact of experience and A.

24:39

I.

24:39

So radiologists with less experience as well as

24:42

radiologists with more experience and benefit

24:46

from interpretation of mammography with A.

24:49

I.

24:50

So.

24:51

I think it's really exciting that we have data

24:53

to support the fact that AI can improve the

24:56

performance of general radiologists to that of

24:59

subspecialized breast imagers and can improve

25:02

the performance of less experienced radiologists

25:05

to that of more experienced radiologists.

25:08

What about interval cancers?

25:10

Those are cancers that are detected between

25:12

screening intervals and they're bad actors.

25:15

They're larger, they're higher stage, more frequently

25:18

ER negative and more, more frequently node positive.

25:22

And so in a paper that was published, um,

25:25

uh, look, looking at interval cancers, they

25:28

looked at the mammograms preceding the cancer

25:30

diagnosis of 429 consecutive screen detected

25:34

cancers, and they had analyzed it with AI.

25:36

Okay.

25:37

And they also have two experienced radiologists

25:39

that classified the prior mammograms, where

25:42

the cancer was as either negative minimal signs

25:45

of cancer, or a false negative so they knew

25:48

where the cancer ultimately was diagnosed.

25:51

And here you can see that 67 percent of the

25:56

radiologists, 67 percent of the prior mammograms

26:00

were either classified as minimal or essentially

26:03

no signs of cancer, but 19 percent of the intro.

26:07

Interval cancers visible as false negatives or

26:10

not identify the signs of cancers by a radiologist

26:14

were classified as the highest AI score.

26:17

So what does that mean, it means that we can find a

26:21

significant number of cancers, a year earlier, when

26:25

we implement AI to, um, to uh, screening mammograms.

26:31

And in a very recently published paper, the largest

26:34

study looking at over 2, 000 interval cancers,

26:38

AI identified the cancers one year earlier in 37.

26:42

5 percent of cancers.

26:43

And interestingly, if you now get more

26:46

complex and add density that they found,

26:51

uh, the 51 percent of the interval cancers.

26:54

So I think what this shows is that we have A very

26:57

powerful tool to deter to identify earlier cancers.

27:02

a year earlier than we're diagnosing them now

27:04

when we, uh, implement AI, when we use AI.

27:08

And if we get more complex using

27:10

density as well, we do even better.

27:13

So AI and interval cancers, we can

27:16

decrease the number of interval cancers.

27:19

It probably will have an impact on survival,

27:22

although we don't have that data yet.

27:24

And when we get more complex and use other

27:26

factors in the First example of that was this

27:29

very recently published paper using breast

27:31

density, the outcomes can be even more robust.

27:36

In Europe, um, in Europe, most countries

27:40

require double reading of mammograms.

27:43

And the question is, can we use AI as the

27:46

second reader instead of the radiologist?

27:48

Because again, human resources

27:50

are the dearest resources.

27:52

So here's a study that looked at AI with 3D

27:55

mammography, um, without AI on the left, with AI

28:00

on the right, and there was an improved performance

28:03

of, um, Mammographic interpretation, and this

28:07

was true both with 2d and 3d mammograms, and the

28:11

accuracy improves AI, like a second pair of eyes,

28:16

and therefore, AI could decrease the workload by

28:19

replacing a second reader in breast screening.

28:23

With non-inferior sensitivity.

28:25

So in a retrospective study using almost

28:28

16,000 3D mammograms, uh, the workflow was

28:32

reduced by 70% when DBT was used in play

28:35

when AI was used instead of a second reader.

28:39

And so the answer is yes, we can use, uh, AI as

28:43

the second reader in those places that require

28:46

a second, uh, mammographer for interpretation.

28:49

So what this does is that it spans the number of

28:52

radiologists available for mammographic interpretation.

28:57

Now what about the more sort of controversial

29:00

idea of standalone AI, what about the fact that.

29:06

What about if we could have mammograms interpreted

29:09

by AI by a computer and never pass a human eyes.

29:13

So, um, again, mammography units are not the rate

29:18

limiting step in much of the emerging and third world.

29:21

Countries, but it's the ability to interpret them.

29:24

So could we perhaps set up a spoken wheel

29:27

system where we place mammography units in

29:30

rural places, have a centralized area, and only

29:34

those small percentage of mammograms that come

29:37

up as with a very high concerning AI score,

29:42

would they then go for a diagnostic mammogram?

29:46

So what, what is the data show on the left?

29:49

You can see the, um, the, the complete

29:53

line, uh, of radiologist performance of

29:56

mammographic interpretation and the AI with

29:59

the, uh, with this, which is the dash line.

30:02

And you can see not only does it

30:04

perform equally, but AI performs better.

30:07

And on the right, if you look by radiologist or.

30:11

The A.

30:11

I.

30:12

Performed better than the majority than

30:14

55 of the radiologists in this study.

30:17

So A.

30:18

I.

30:19

A standalone AI really can

30:21

function like a radiologist alone.

30:25

And again, we will continue to improve the

30:27

performance that perhaps, and perhaps we're there

30:30

now, that AI performs even better than a radiologist,

30:35

even better than a subspecialized radiologist.

30:39

And in the study by Emily comment, this also was

30:42

shown that the AI performed as well there was

30:46

no difference of the average of the readers and

30:49

the performance in AI and screening mammograms.

30:52

So, standalone AI.

30:54

Yes, it.

30:56

equal to and maybe, you know, better than both

30:59

the general radiologists and the breast imager.

31:02

And it certainly addresses some of the resources with

31:05

regard to, um, the limited resource of radiologists.

31:10

And I really hope that we can think about

31:13

creative ways of spoken wheel system in

31:15

emerging countries to improve, um, the

31:18

availability of life saving mammograms.

31:22

What about false positive calcification detection?

31:25

So there are places that are implementing AI and some

31:28

of the things that people have said is, you know, gee,

31:32

um, when we implement AI, so many of the calcifications

31:36

are noted that we're finding that at the initiation

31:39

of implementation or biopsy more calcifications.

31:43

But here you can see one AI product that, um,

31:47

for calcifications That are below seven, the

31:51

likelihood of malignancy is nine, is less.

31:56

of benignity is 99.

31:57

97 percent.

31:59

So what about if you just never biopsied, no matter

32:02

what you thought, you never biopsied any of the

32:05

calcifications that were seven or below or six and

32:10

below, and with that you can confidently Not biopsy

32:14

classifications that you otherwise might, um, biopsy.

32:18

So, uh, it can, um, really allow us to

32:23

triage the cases based on the the A.

32:26

I score to reduce the number of

32:29

classification with biopsies.

32:33

Now, we talked a lot about, um, uh,

32:38

AI, uh, with the radiologist, 2D, 3D,

32:42

AI alone, the decrease in the time.

32:45

What about risk assessment?

32:46

We know that risk assessment, breast cancer

32:49

risk assessment is becoming much more complex.

32:54

And can we use even more information that

32:57

AI has, you know, and with AI, we know that

33:01

we can glean an enormous amount of genomic

33:04

information from an individual woman's mammogram.

33:08

So even when we use the, um, Different risk

33:12

assessment tools tire Q six gale model.

33:16

There's still, um, population based, you know, if

33:19

a woman has a has a D H and the family history,

33:24

it's still based on population based information.

33:27

But what about if we could glean the specific

33:30

genomic information in a woman's mammogram,

33:33

so we can determine her risk assessment, and

33:37

there is preliminary data to suggest that AI.

33:41

Can do exactly that.

33:42

And maybe we can even further, um, stratify risk based

33:48

stream is screening on a woman's genomic information.

33:52

And in that regard, We could further have

33:57

further availability of screening mammography

34:00

and those resources, particularly in emerging

34:03

countries that don't have any screening to

34:06

maybe impact those unacceptable mortality rates

34:10

from breast cancer, and we are beginning to see

34:13

that we saw today I shared with you one of the

34:16

studies that looked at implementing not only AI.

34:20

But AI with breast density with

34:22

improved outcome of identifying and

34:26

diagnosing, um, interval breast cancers.

34:29

So, um, artificial intelligence and

34:33

mammography, can it level the playing field?

34:36

It can increase the performance of a general

34:39

radiologist to that of a breast imager.

34:42

It increased the performance of a less experienced

34:45

radiologist to a more experienced radiologist.

34:48

It can, it can, Potentially wipe out the

34:52

disparity geographic disparity in interpretation

34:56

differences and allow for interpretation of

35:00

the highest quality in socioeconomic areas that

35:05

might not have access to centers of excellence.

35:09

It could definitely work with both

35:10

2d and 3d mammography, and it's

35:14

highly effective as a second reader.

35:17

The ability to Detect interval cancers early

35:20

can substantially and meaningfully likely impact

35:25

survival as well as costs, and it's highly effective

35:29

in standalone interpretation and in a screening

35:32

population has a negative predictive value of 99.7%.

35:37

So what about the idea that human eyes will never

35:40

see 70, 80, meaning maybe 90% of mammograms globally?

35:45

And yet we could accurately interpret mammography.

35:49

And this could be especially true in those parts of

35:52

the world that have markedly limited human resources.

35:56

So one of the things that I think,

35:58

what's the elephant in the room?

35:59

So many people, um, discuss whether

36:02

AI will displace radiologists.

36:04

And the answer is no.

36:06

And my hope is that we won't hinder implementation

36:10

of AI because of the fear of displacing radiologists.

36:15

AI will help healthcare professionals,

36:18

not displace healthcare professionals.

36:21

It'll give them the time and the tools that

36:23

they need to focus on what matters and what

36:26

needs true human decision making to build a more

36:30

efficient and intelligent system for patient care.

36:34

And it's here and now studies have shown that you

36:37

know 70 percent of practices, radiology practices

36:41

plan on implementing AI in the near future.

36:44

And really, AI is here and now and the question is,

36:48

how do we optimally integrate AI into our clinical

36:51

practice to improve our care, and to improve the

36:55

accuracy of our interpretation and our productivity,

36:59

and it's not whether we will the reality is, we will.

37:03

Um, it's just how will we how quickly will we and

37:07

how will we best situate the implementation of AI

37:11

in our practices, um, to achieve the best that we

37:15

can for our, for our patients and for ourselves.

37:18

And so can AI really disrupt the

37:21

disparities in breast cancer.

37:24

And the answer is yes it's

37:26

it's a remarkable contribution.

37:30

incredible potential to really impact

37:34

the plane, the impact disparities,

37:37

both in the United States and globally.

37:40

And so with that, I thank you.

37:43

This is my, um, this is my email.

37:46

Please feel free to email me.

37:49

Um, and, um, I think now we certainly have the

37:52

opportunity for some questions and answers.

37:56

So, um, see.

38:01

Let's see.

38:03

Trying to find the questions here.

38:07

Young generation of radios that AI is an ally of next

38:10

generation radiologists and not a rival that will steal

38:13

our job so I think I think it was more confirmation of.

38:20

of what we've discussed today.

38:22

Let's see about the questions and answers.

38:25

Um, so there's a question that somebody

38:28

asked, uh, is what tracers are being

38:30

investigated for molecular imaging?

38:33

Um,

38:35

a question is, do most breast imagers

38:37

believe that ROC of diagnosis is better for

38:40

AI alone compared to any group of imagers?

38:44

I'm not sure I understand, but I think what

38:46

the data shows is that With the improving,

38:49

um, uh, with the improving performance,

38:55

um, of that the ROC demonstrates.

38:59

Um, AI alone, the, the question of AI

39:05

alone versus any group of imagers are

39:07

still, is still a, an area of research.

39:10

Um, and it also depends on what, what

39:14

version of AI we're talking about.

39:16

So we believe that we might get to a point where AI

39:20

alone could handle screening mammography, and therefore

39:25

radiologists could really focus on diagnostic.

39:27

Evaluation and biopsies.

39:30

But I think that's the question.

39:32

And right now we are absolutely not at a position

39:37

in this country or anywhere where a I is going

39:41

to replace interpretation of radiologists.

39:44

But I think the hope is that we will be able

39:46

to perhaps implement that sooner in parts

39:50

of the world that have a worse prognosis.

39:55

And have no screening.

39:57

Um, wondering if there's data

40:00

about uptake of a I globally.

40:03

Um, I know, you know, I can't answer that right now.

40:06

I don't have the numbers, but I do know that

40:09

a I is being implemented around the globe.

40:13

Um, in, in various areas.

40:15

I know, uh, that with one AI product, it's in

40:19

over 30 countries now and it with transparent,

40:22

it's in over 30 countries across the globe now.

40:25

And, um, and so, and I'm sure that's

40:29

true for other AI products as well.

40:32

And then, um, How do I address the following but

40:37

it doesn't ask what that what the following is so

40:41

perhaps I could hand this back over to MRI online.

40:45

How do I know,

40:48

how do I know AI could improve mammographic

40:51

acquisition so we didn't go into that but there is.

40:55

Um, There is a I even available a I products

40:59

now that review a mammogram immediately after

41:03

it's acquired and will give feedback to the

41:06

technologists for positioning and technique.

41:10

So, um, so that's the answer to that.

41:16

The next question is most of the

41:17

studies are retrospective and small

41:19

numbers from selected populations.

41:22

So will this affect performance in other populations?

41:24

So I think we know that we will need to

41:29

include populations that we want to use AI for.

41:33

And I think that's a really good point.

41:35

But I, I also believe that, uh, there are

41:38

quite a few, uh, studies out of Europe.

41:41

There are quite a few studies out of the United States.

41:44

The performance is fairly similar.

41:47

Um, even though there, there may be

41:49

differences in algorithms and populations,

41:52

but I do think it is important.

41:54

to do studies in Africa before we say that it's equal

41:59

to the performance in Europe and the United States.

42:01

However, the need is so great that even if the

42:04

performance, and I'm not saying it is, but even

42:07

if the performance is slightly lower in Africa,

42:11

it would impact this really dire, um, survival

42:16

rate, uh, in your, in Africa that we see today.

42:22

Um, so I don't see Any other questions

42:27

and perhaps I can hand this back.

42:29

Let me see one more here.

42:32

Uh, okay.

42:34

So I can hand this back to MRI online.

42:37

I want to thank everybody for, um,

42:39

joining me today, joining MRI online.

42:43

And, um, and if you'd like to reach

42:45

out to me again, this is my email.

42:48

Thank you so much, Dr.

42:49

Brim, for that fantastic presentation.

42:51

As we bring our time together

42:52

to reclose, I want to thank Dr.

42:54

Brim for this lecture, and thanks to all

42:55

for participating in our new conference.

42:57

A reminder that you can access the recording of

42:59

today's conference and all other previous student

43:01

conferences by creating a free MRI Online account.

43:04

If you'd like to access our Mastery Series courses,

43:06

Case Series, and much more educational content with

43:09

unlimited CME, you can sign up for a free seven

43:11

day trial of our MRI Online Premium Membership.

43:14

You can learn more at MRI Online

43:16

at the URL in the chat box.

43:18

Be sure to join us next week

43:19

on Tuesday, March 22nd at 12 p.

43:21

m.

43:22

Eastern Time for a lecture with Dr.

43:23

Lisa Ann Mullen on MRI Basics and Virads.

43:27

You can register for that lecture at mrinline.

43:29

com and follow us on social media at The MRI Online

43:33

for updates and reminders on upcoming noon conferences.

43:36

Thanks again and have a wonderful day.