0:01

Hello and welcome to noon conference hosted by MRI

0:04

online. Noon conference was created when the pandemic hit

0:07

as a way to connect the global Radiology community Through free

0:10

live educational conferences that are accessible for all.

0:13

It has become an amazing weekly opportunity to learn alongside

0:16

Radiologists from all around the world and we encourage you to

0:19

ask questions and share ideas to help the community Learn and Grow.

0:22

You can access the recording of today's conference in previous

0:25

new conferences by creating a free MRI online account.

0:28

The link will be provided in the chat box. You can also

0:31

sign up for a free trial of MRI online premium membership to

0:34

get access to hundreds of case-based microlearning courses

0:37

across all key Radiology such specialties.

0:40

learn more at mrionline.com

0:43

Today we're honored to welcome Dr. Susie bash for a

0:46

lecture on AI applications in advanced neuroimaging.

0:50

Dr. Bash is the medical director of neuroradiology at San

0:53

Fernando Valley Interventional Radiology at radnet.

0:56

Prior to this. She was assistant professor of

0:59

neuro radiology at UCLA.

1:01

Dr. Bashas' passions and interests lie in

1:04

artificial intelligent applications in Advanced neura Imaging

1:07

serving as a consultant to several AI companies. She's also

1:10

a member of the editorial board in the AI section

1:13

of Applied radiology and is actively involved in

1:16

AI clinical trials and peer-reviewed Publications as

1:19

well. She is recurring guests on TV radio and

1:22

podcasts.

1:24

At the end of the lecture join Dr. Bash in a Q&A session

1:27

where she will address any questions you may have on today's topic.

1:31

Please remember to use the Q&A feature to submit your questions we can

1:34

get to as many as we can before our time is up.

1:37

But that we are ready to begin today's lecture Dr. Bash.

1:40

Please take it from here.

1:42

Hi, my name is Susie bash. I'm a neuro radiologist

1:45

at Redneck. Thank you for joining us today for the MRI

1:48

online noon conference. Today. We're going to be discussing AI applications

1:51

in advance Neuro Imaging.

1:55

So these are a few disclosures.

1:59

So AI Solutions are playing a very important role

2:02

in neuroimaging. It's important

2:05

to sort of understand what tools exist and which may

2:08

be best applicable for your clinical practice to enhance

2:11

patient care.

2:12

Artificial intelligence is a computer

2:15

method that performs tasks typically requiring human

2:18

intelligence machine learning is a subset

2:21

of AI which enables computers to learn from existing data

2:24

without explicit programming. So machine learning

2:27

can be further subdivided into unsupervised and supervised learning

2:30

for unsupervised learning the computer itself

2:33

determines the image class but in supervised learning

2:36

some ground truth exist to train the

2:39

algorithms so deep learning is typically supervised learning

2:42

method that uses some form of

2:45

neural network. Typically a convolutional neural network

2:48

for automated image classification. And

2:51

I think one of the fascinating things about deep

2:54

learning is these algorithms can actually learn from

2:57

their mistakes and improve over time when

3:00

fed more data.

3:02

So AI applications can be applied before during

3:05

or after image acquisition.

3:08

Some of the AA applications I can be performed

3:11

before image acquisition would be scheduling Insurance authorization

3:14

billing mining packs

3:17

and medical records and help with no shows.

3:20

So I work for the largest outpatient Imaging Enterprise

3:23

freestanding Imaging Enterprise in the US and we

3:26

do about eight million exams a year about 350 imaging

3:29

centers and about 350 Radiologists. So

3:32

we actually either incorporate a lot

3:35

of AI companies and tools in our practice or we partner

3:38

with them. So in our practice we do use AI for

3:41

helping us with scheduling particularly for mammograms.

3:44

We use AI for help in Billing

3:47

and also to help prevent no shows

3:52

So AI applications that you can apply during image acquisition

3:55

would be things like to help optimize image acquisition

3:58

deep learning reconstruction and

4:01

even synthetic Imaging.

4:04

So let's start with optimization of Imaging angulation. This

4:07

is an AI tool called CT copilot from

4:10

cortex. You can see on the

4:13

top row here. This is before any solution was

4:16

applied afterwards. You can see how you can

4:19

angulate the brain perfectly with the AI tool it

4:22

also performs automated segmentation and

4:25

automated subtraction match. So

4:28

you see here the current study. This is the prior study and

4:31

this image on the right is really a heat map overlay

4:34

of the current and prior so you can see how much

4:37

larger these ventricles have gotten over time, which can be quite useful

4:40

in the setting of hydrocephalus.

4:45

This is a product from Siemens. They can

4:48

help automatically label spine studies.

4:51

It can label the ribs for you and

4:54

also it can sort of splay the

4:57

ribs out which it can be very helpful to detecting

5:00

bone fractures or RCs metastases.

5:03

Now another tool that can be applied during

5:06

image acquisition is deep learning reconstruction. I actually

5:09

think deep learning is one of the most exciting and

5:12

revolutionary things that has come out in a

5:15

time in the 22 or so years that I've been practicing patients love

5:19

it. It allows us to go much faster

5:22

and get patients in and out of the scanner and you know

5:25

60 to 70% faster it the

5:28

Radiologists love it because he can improve image quality and

5:31

imaging Enterprises also love it is an

5:34

improves workflow efficiency. So basically a DLR

5:37

allows Superior perceived image quality higher

5:40

perceived signal to noise ratio higher perceived spatial

5:43

resolution higher perceived contrast and always

5:46

ratio reduced artifacts and reduce dose

5:49

for example and PET CT and it enables faster

5:52

scans.

5:54

So how does DLR work? Well Imaging facilities

5:57

alter their protocols to decrease the scan time,

6:00

which I call fast scans and that acceleration can

6:03

be accomplished either by understanding and reducing excitations or

6:06

by reducing the image matrices.

6:09

Then DLR is applied to the fast data

6:12

set to restore image quality by denoising which

6:15

you use for both Brain and Spine and other areas throughout the

6:18

body and also sharpness enhancement, which can be

6:21

applied in the brain.

6:23

Now they're a vendor neutral solution that

6:26

offers DLR is subtle medical. They were actually first to Market

6:29

space and then the oems all the oems now have

6:32

some DLR product at different

6:35

stages of fruition of for FDA approval but most are now

6:38

FDA approved. Gee has air recon DL Siemens

6:41

has deep resolve Cannon has advanced intelligent clear IQ

6:44

engine Phillips has smart speed

6:47

and FDA pending Precise Imaging and Fuji has

6:50

IP rapid.

6:52

So which one you go with that vendor neutral solution or

6:55

an oem you may want to use both in your clinical

6:58

practice. We actually use both but in advantage of

7:01

the vendor and neutral solution like subtle medical is they

7:04

operate in the dicom space rather than in case

7:07

based like the oems so they can be applied to any scanner brand

7:10

of any age. So really provides

7:13

a virtual upgrade to Legacy scanners,

7:16

which can include the ROI by extending the

7:19

scanner life and preventing having to do a hard forklift upgrade

7:22

or replacement. OEM dll

7:25

products are really kind of limited to their newest

7:28

scanners at this point in time. So we

7:31

tend to use an oem Solution on our newer scanners

7:34

and settle Medical on our older scanners, but we

7:37

but subtle medical. We also use on some of our newer scanners

7:40

as well.

7:42

So this can take a routine brain MRI exam slot

7:45

from 30 to 40 minutes down to essentially 15 minutes

7:48

and that's not to scan time that's on the table and

7:51

off the table time in addition to scan time.

7:55

This is what it looks like. Here's a subtle medical here subtle Mr.

7:58

Here's the standard image on the left the fast image the

8:01

noisy image in the middle. And then if

8:04

you take this noise image and apply deep learning. This is what you

8:07

get here. But at half the scan time

8:10

you see a bump here in the signal to noise ratio and overall image

8:13

quality again standard here fast here.

8:16

And this is after you apply deep learning for reconstruction standard

8:19

here the fast one here

8:22

and look here at the image quality. It's just

8:25

it's actually much better than what even the standard of care is. This is

8:28

a nice example because it demonstrates how DLR improves

8:31

contrast and noise ratio. This is a little cortically

8:34

based metastasis here. It's very difficult to see there's

8:37

another one here, but you see how it's hard to see on the standard of

8:40

care image when you accelerate your protocol and go faster and

8:43

you get a bump and then contrast to noise ratio. Now you can see

8:46

it better, but the image looks noisy you then

8:49

take this noisy image apply deep learning and now you maintain that

8:52

high contrast to noise ratio, but you also,

8:55

Improve your signal to noise ratio. And this is just you

8:58

can see how detecting these small metastases could

9:01

make a difference and save a patient's life. So this is

9:04

a very nice solution. It also decreases motion

9:07

artifact your standard of care and here is DLR

9:10

but yet 72% faster much less

9:13

motion appearance on this image. Here is

9:16

GE or Recon DL. This is a

9:19

standard of care here. Look at the gray white Junction and

9:22

look how much that is improved after we applied deep learning.

9:26

Here's Canon's Ace product their standard

9:29

of care. This is after you apply deep learning. Just

9:32

this is just a significantly better image much

9:35

sharper much higher SNR. Here's another

9:38

example from Canon. Look at the hippocampus here on the standard

9:41

and after deep learning is applied. This is GE here

9:44

images on the left or the original images on the right or after

9:47

we apply DL another GE one here, the

9:50

original and after DL is applied. This is

9:53

medic Vision. This one actually is the only example that's not

9:56

people learning. This is machine learning. But again, we got

9:59

some acceleration 22% faster in improved image

10:02

resolution. This is a deep

10:05

learning product from subtle, Mr. Called subtle pet. We were

10:08

able to go take an exam slot for this

10:11

PET CT from 20 minutes down to five minutes 75% faster

10:14

apply the Deep learning to restore the

10:17

image quality, but at significantly less radiation dose,

10:21

This is another product that this one's actually in

10:24

development by the way, all the products. I mentioned will be

10:27

FDA approved unless I specifically say on my slide that it's in

10:30

development. This is called subtle Gad. This is a hundred

10:33

percent contrast here in this brain tumor. If you

10:36

drop that contrast dose down to 10% you lose

10:39

any visual enhancement of the brain tumor, but if you

10:42

apply deep learning to this 10% contrast dose,

10:45

you can restore the appearance of contrast enhancement.

10:49

And this is actually very pertinent in this

10:52

day and age when there's so much talk about gadolinium deposition throughout

10:55

the body and including in the brain. So we're

10:58

hoping that this can you know, come to

11:01

clearance soon. It's a very exciting product now, ai

11:04

another application during image acquisition is

11:07

synthetic Imaging. This is a

11:10

product that's in development that's happens to be subtle synthetic Imaging

11:13

and basically what we're doing is creating a

11:16

stir image from a T1 and T2 input

11:19

so here have here.

11:21

Actual stir image here. And this is the synthetic

11:24

stir my to my eye. I

11:27

actually prefer the appearance of the synthetic to the actual scan.

11:30

Here's the actual scan here on the left and the

11:33

synthetic image on the right again actual

11:36

scan on the left synthetic image on the right to me

11:39

this the signal to noise ratio of appears higher

11:42

on the synthetic image and obviously, you know stir can

11:45

take time so you can really cut a lot of time off your exam if

11:48

you use this synthetic acquisition here is

11:51

the T1 input at the t2 input and

11:54

then from that we've created the synthetic image here

11:57

on the right. This is the actual stir this is a synthetic image

12:00

again here the T1 input T2

12:03

input in this patient with this Titus and osteomyelitis.

12:06

This is the actual stir and this is the synthesizer

12:09

on the far right again actual

12:12

scan on the left synthetic synthetic stir

12:15

on the right actual on the left and

12:18

the synthetic on the right. Look at the difference here in the quality the Sun.

12:21

The synthetic image actually just looks like it

12:24

has overall better quality.

12:26

Now there are a lot of AI applications that can be applied after

12:29

image acquisition. This is not an inclusive list, but

12:32

these are some of the key products that are out on the market space

12:35

today AI is actually showing tremendous promise

12:38

in cancer screening. I'm not

12:41

going to focus really too much on anything outside of the brain.

12:44

But at radnet we do use quantity that

12:47

can help us detect prostate cancer and also segment the

12:50

tumors. We all also partner with Ezra again

12:53

can detect prostate cancer and segment the tumors. We

12:56

use Incorporated a company deep Health.

12:59

They showed in a nature article that

13:02

their AI tool can detect breast

13:05

cancer on mammograms one to two years before expert Radiologists.

13:08

This is just very very exciting application

13:11

and can really help save lives.

13:14

We also Incorporated a company aidens, which is in development

13:17

that can detect lung cancer and another

13:20

partner another AI company that

13:23

we partner with is cortex. They have an On-Q neur.

13:26

Product that detects brain cancer and segments the brain

13:29

cancer. We'll talk more about that shortly.

13:32

So another AI application during image acquisition would

13:35

be triage apps. There are a lot of triage apps

13:38

out on the market space and these are very exciting particularly

13:41

utilized in the hospital-based setting

13:44

it can detect critical findings on Imaging and

13:47

prioritize those studies. So the top of the Radiologists work

13:50

list so that the patient can be treated earlier. This

13:53

is adox product that's detecting a cervical

13:56

spine fracture and flagging it. You can see how something like

13:59

this might be very helpful to a new Resident that's

14:02

you know doing an ER shift and having that

14:05

second look from the AI tool. This is actually applied

14:08

even before the you know, the radiologist sees it

14:11

but can really be helpful and getting

14:14

the patient treated earlier. This is a exciting

14:17

product from visai. This is their lvo large

14:20

vessel inclusion detection algorithm. It basically

14:23

stands the image

14:26

before even the radiologist sees it it will

14:29

detect the lvo and send via a HIPAA secured.

14:32

Cell phone app to everyone in the in the

14:35

network for the stroke team.

14:38

So here we get AI powered notifications

14:41

for both the lvo and for perfusion High

14:44

Fidelity mobile image viewing real-time patient

14:47

information and full stack secure

14:50

communication. So you can actually chat between the members of The

14:53

Stroke team and in this really a is has

14:56

become extremely useful for stroke

14:59

work up. This is again an

15:02

example. This is what they're CT perfusion algorithm looks

15:05

like on the cell phone and basically

15:08

it operates immediately after

15:11

the scan is taking 63 seconds from

15:14

the scan time to alert of the entire multi-disciplinary stroke

15:17

team. And this is really proven

15:20

to save time and they come in Continuum of

15:23

Care 102 minutes saved

15:26

in the door in and door out 52

15:29

minutes faster than standard of care from the

15:32

time.

15:32

CT was obtained to the

15:35

time of notification of the team

15:38

and treatment 86 minutes saved from

15:41

door into a growing puncture for mechanical thrombectomy.

15:45

This is just an example again of the

15:48

the time savings here 73% faster

15:51

before team notification 25 24%

15:54

faster door to intervention. This is

15:57

utilizing over 900 hospitals in the US right now

16:00

and it really increases access to

16:03

care. So instead of the patient being routed to a hospital

16:06

that's not enabled for mechanical thrombectomy. If

16:09

you start early here, you can get the patient to the

16:12

correct Hospital faster. It also decreases

16:15

day 2.5 days in

16:18

hospital stay and it also improves clinical

16:21

function.

16:24

This is again showing 3.5 days are saved

16:27

in neural ICU. This is length of stay in days

16:30

2.5 days in hospital stay and it

16:33

also amounts to a savings for patient expenses as

16:36

well. So, you know, if a potential a length

16:39

of stay for an untreated ldo might be 15 days at

16:42

$3,000 per day 45,000 total

16:45

cost of stay but CMS will

16:48

only maybe reimburse for 13,000. So that

16:51

patient in you know has a loss of 31,000 666

16:54

but again,

16:57

the more important issue than the finances is

17:00

actually clinical outcome differences and this makes a

17:03

big difference in clinical outcome. In fact

17:06

so much of a difference that CMS actually provided, you

17:09

know, a new technology reimbursement package.

17:13

So if hospitals use this there is a 1,044 reimbursement

17:17

because it makes a huge difference in patients' lives

17:20

here is a rapid product AI solution.

17:24

That's again detecting the large vessel effusion. Here's

17:27

again showing the CTA this

17:30

Red Zone here. It

17:33

shows the blood vessel density on this

17:36

scale here. So red is significant decrease in

17:39

density. This is rapid ai's perfusion tool

17:43

that shows the mismatch here the mismatch

17:46

of 63 milliliters and a mismatched ratio

17:49

of 2.3 between the ischemic core and the

17:52

perfusion deficit the penumbra here. Here's again,

17:55

this one is actually the Mr. Profusion tool

17:58

that rapid has ischemic pore and perfusion deficit

18:01

with a mismatch. This is the aspect scale

18:04

that the tool that

18:07

rapid AI uses and this is very important because typically

18:10

patients aren't mechanic candidate for mechanical from

18:13

that from back to me if they're aspects for

18:16

is less than four. Now. There are a lot

18:19

of triage apps out on the market space that can detect intercranial Hemorrhage

18:22

automatically.

18:24

Flag those and and then you can get prioritization. This is

18:27

zebra medical tool. Here's AI doc that's

18:30

detecting this subdural here. Here's

18:33

Max Q AI.

18:35

This is Rapids detection tool here

18:38

where you see the prank highlighted in

18:41

red this tool. I really like this is icometrics

18:44

tool. It does not only detects the subdural hematoma

18:47

is also detects blood in other areas throughout

18:50

the brain but it also gives you midline shift

18:53

and cisternal compression. You actually get a whole report here. This

18:56

is their TBI report. And so this

18:59

is a very useful tool because some of a lot of

19:02

the AI tools will only just detect the blood but not give you

19:05

the volume of the blood and there's a lot of subjectivity

19:08

when it comes to measuring things like subdural hematomas

19:11

to the concavity of the calvarium. So it's nice

19:14

to have tools that actually calculate the volume busy. I

19:17

also has a tool that calculates the volume. These are some other

19:20

tools from viz. They've got the lvos we discuss they have

19:23

aneurysm detection into cranial hammerich detection

19:26

and subdural detection. Rapid AI

19:29

here has an intracranial aneurysm detection tool. It

19:32

gives the pertinent measurements for the aneurysm. So that's

19:35

a

19:35

Full product synaptive here

19:38

their AI tool does tracking of

19:41

white matter tracts. This can be used interoperatively in

19:44

real time as the neurosurgeon is

19:47

guiding the needle you'll know exactly where he

19:50

is.

19:51

Now in AI solution that can be applied after image acquisition

19:54

is quantitative volumetric Imaging. This

19:57

is been we've been I've been using quantitative volumetrics

20:00

and my clinical practice for over 15 years. Now

20:03

this basically identifies and

20:06

labels anatomic structures and then quantifies

20:09

the volume of those brain structures and Compares that

20:12

to an age and gender match normative database and this provides

20:15

for volumetric tracking to assess for rate of change

20:18

over time. Now, there are several different companies that

20:21

are in the Quant space. This is a cortex AI

20:24

they were actually first to Market space and this is the one that I've

20:27

been using for 15 years my clinical practice. This is

20:30

I go metrics. He had a look here at what their

20:33

segmentation looks like. This is a quantum

20:37

And this is cominastics this

20:40

I'll throw in here. Although this product

20:43

is slightly different from the other Quant tools that I'll

20:46

be talking about this actually that does a darn

20:49

darmian's tool here does

20:52

voxel analysis from the MR input. So

20:55

sort of microscopic analysis. They combine that with

20:58

clinical biomarkers to try to

21:01

predict which patients will progress from mild cognitive

21:04

impairment to Alzheimer's disease that you

21:07

have FDA breakthrough status, but this product is in development Quant

21:10

can also be used to segment individual

21:14

lesions in the brain as you see here in this example from cortex.

21:17

And so why is quantitative Neuro Imaging

21:20

appealing? Well, it improves diagnostic value

21:23

it eliminates report bias, which by the way happens

21:26

all the time and our refers really they don't

21:29

like report by it. You know, some people will say moderate

21:32

cerebralette fee commencement with age and then they'll come

21:35

back a month later and

21:37

Else will say, you know moderately severe cerebral

21:40

atrophy with a temporal parietal predilection and the referring

21:43

doesn't know should I be worried about Alzheimer's or is this just

21:46

normal for age? And so this really helps in that

21:49

it's easy to add on and negligible cost acquisition speed.

21:52

No external hardware is required. It's cloud-based.

21:55

It integrates seamlessly and quickly into the

21:58

packs, you know, really processing in just

22:01

five to seven minutes no Radiology post-processing to

22:04

required. It's easy to interpret and it actually offers a

22:07

big referral Vantage our refers that use Quant. I've never

22:10

had any stop using it once they started using it. These are

22:13

just the protocol requirements. So for if

22:16

you're doing a dementia, you really only absolutely need

22:19

sort of a team one. We prefer 3D thin size

22:22

acquisition at one millimeter for Ms. You're going to want

22:25

to have flare two again. We prefer 3D acquisition and

22:28

these are sort of some of the different

22:31

clinical indications that you can use Quant for so

22:34

there's really a lot of them. We actually on our

22:37

Form, we have a click box for quantitative volumetric Imaging and

22:40

the referred can actually click what kind of protocol they

22:43

want to use the technologists then

22:46

sends the input sequence through the

22:49

cloud portal to the company and the

22:52

a report is generated based

22:55

on the clinical indication. So let's look now at

22:58

dementia. This is one of I would say of all these indications. We probably use

23:01

Font most for Dimension and multiple sclerosis. So there

23:04

are a lot of different causes of dementia.

23:08

But we often talk about Alzheimer's disease because this

23:11

makes up the largest population by autopsy series

23:14

six million Americans are suffering from Alzheimer's disease.

23:17

It doubles in frequency really every five

23:20

years after the age of 60. So last year

23:23

across our nation 355 billion dollars one in

23:26

three of our seniors will die of dementia. They

23:29

really kills more people than breast cancer and prostate

23:32

cancer combined and just for you know, for example

23:35

that death from heart disease has gone down

23:38

by seven percent since the year 2000 but from

23:41

Alzheimer's disease, it's gone up by a hundred and forty-five percent. So

23:44

this is a really big Health population issue. It's characterized

23:47

by Beta amyloid plaques which are extracellular and

23:50

can be image by amyloid pet and by neurofibrillary Tangles

23:53

that are intracellular towel and they can be

23:56

image five tall pet. This is an amyloid pet. This is a towel

23:59

pet animal pets actually positive preclinical stage

24:02

of Alzheimer's disease even up to 20 years before the patient's symptomatic

24:05

and then amyloid plaque deposition.

24:07

Eye disease we see a lot of it in

24:10

Alzheimer's disease. We can see some in-dimension with Lewy

24:13

Body and you shouldn't really see any with front or temporal dementia.

24:16

So, you know again the patient comes in for cognitive

24:19

testing and MRI is often ordered in really

24:22

a large part to exclude any other pathology that

24:25

might be causing memory loss. We always encourage our

24:28

refers to order quantitative volumetric Imaging to attack

24:31

that onto the MRI when they come in and then if there remains

24:34

clinical ambiguity a pet can be ordered we

24:37

have fdg pet amyloid pet and pal pet amyloid and

24:40

Tau are not reimbursable right now by CMS.

24:43

So most people don't tend to order

24:46

it because it's a large amount of money out of pocket. I've read over

24:49

200, you know amyloid pets in

24:52

my career just because I it was a reader for

24:55

the ideas trial but I don't see nearly as many ordered now that

24:58

you know, since they're not reimbursed. Hopefully that

25:01

will change in the near future particularly since we're hoping

25:04

for disease modifying therapies that

25:07

are kind of

25:08

Now this is a pet Mr. Fusion here where you see the

25:11

critical hypo metabolism. This is PET

25:14

CT Fusion cortical hypo metabolism in the temporal lobes and

25:17

again a color-coded PET CT Fusion. This is what

25:20

a positive amyloid pet looks like where you see diffuse binding of

25:23

the Tracer throughout the cortex here and this is

25:26

the color map overlay just for comparison. This is

25:29

what a negative pet looks like where you see this sort of tree and Branch appearance and

25:32

really don't see any binding here in the cortex.

25:36

So let's look at a couple cases of how we

25:39

apply font. This is a 79 year old with memory loss.

25:42

This was the original MRI in

25:45

2013. And we see some cerebral atrophy here just

25:48

sort of mild to moderate in degree in the right temporal lobe

25:51

but the hippocampal volumes were still okay, the

25:54

ventricles are actually slightly enlarged outside

25:57

of the mean here anything in this team

26:00

zone is outside of the mean this happens to be neural Quant.

26:03

This is a cortex product. This is Echo brain, which is an icometrics

26:06

product and then the patient came back three

26:09

years later and we see significant progression and meso

26:12

terminal atrophy. And and then we see here

26:15

a drop off the normative curve for the particular

26:18

pertinent biomarkers such as the hippocampal

26:21

volumes and the hoc and we

26:24

also see significant enlargement of those ventricles as the hippocampi atrophy

26:27

you get compensatory enlargement of the

26:30

inferior lateral ventricles. This is the neuro Quant

26:33

the Ico grade and this is a heat map overlay from the

26:36

Prior study point to the current study. This

26:39

is again a closer. Look here. They have to

26:42

cancel volumes have dropped off the normative curve where outside of two

26:45

standard deviations from the mean hippocampal occupancy is

26:48

also dropped an inferior lateral ventricular volume

26:51

has gone up you see here. We're in the 99 percentile. This

26:54

is the icobrain report again less than

26:57

1% tile on both studies. But

27:00

now the temporal cortex has become statistically

27:03

significant here at less than 1% on the

27:06

follow-up study. This Apples by graph from

27:09

the Ico metrics report here where it's pointing

27:12

to areas of statistical significance. The blue is the

27:15

one percent one percentile 10 is

27:18

I'm sorry. The Orange is the 10 percentile. You

27:21

can see here with segmentation that those hippocampi have

27:24

atrophied in time over the three years. This

27:27

is neuroquence heat map overlay anything

27:30

in red is getting bigger. So the ventricles are

27:33

getting bigger and the souls are getting bigger anything in

27:36

blue.

27:36

Is getting smaller so the cortex is shrinking over

27:39

time. So this is a very nice visualization of the current superimposed

27:42

on the prior. And this heat map this

27:45

patient just incidentally also had a happen to have

27:48

cerebral and angiopathy you see little fociable Hemisphere

27:51

and stain insteaded throughout the brain the patient then

27:54

went on to have an amyloid pet and this is diffusive positive

27:57

diffuse finding throughout the cortex. So this patient had Alzheimer's

28:00

disease and also it's regal amyloid angiopathy.

28:03

The next patient is an 86 year old with memory loss.

28:06

Here's the initial MRI just mild strugglect

28:09

through the right Temple predilections the

28:12

little Lacuna infarct there in the ponds the fgt pet

28:15

showed hypo metabolism here in the

28:18

bilateral Museum temporal lobes. This is the PET CT Fusion.

28:21

Here's the pet Mr. Fusion. This was statistically significant

28:24

when I ran it through the memory analysis software. I

28:27

run all pet studies whether amyloid or fdg through

28:30

memory analysis, which tells you kind of what is

28:33

statistically significant for patient age, which is

28:36

Really helpful tool you can see the patient came

28:39

back in 2016. Now we have really relatively severe at

28:42

Via the temporal lobes and statistical significance on

28:45

quantitative volumetric Imaging. Here's a

28:48

look at that neural Quant report and we're way outside of

28:51

normal here in the Pink Zone hippocampi are

28:54

at one percent. The inferior lateral ventricular volumes

28:57

are at 99% And by the way, you may be

29:00

wondering what the hippocampal occupancy score is. It's basically

29:03

an estimate of the degree of hippocampal atrophy. So it's

29:06

the hippocampal volume of the hippocampal plus inferior lateral

29:09

ventricular volume on the left side plus the

29:12

right side and averaged here. And so the lower

29:15

hoc scores are highly associated with progression from

29:18

mild cognitive impairment to Alzheimer's disease. So the

29:21

bottom line is the lower hoc, the more at risk

29:24

the patient is for developing Alzheimer's neuroquan also

29:27

has a triage brain at tree report that's divided

29:30

into the different lobes with substructures anything in the

29:33

red is more than two standard deviations outside of the

29:36

knee.

29:37

Here's a look here. This is from the iGo brain report

29:40

that's showing that everything is really statistically significant

29:43

by the time they came back in 2016.

29:47

There's so this patient went on to have an amyloid PET CT. This was positive. This

29:50

patient has Alzheimer's disease. Next patients

29:53

is 73 year old with memory loss. Here's the

29:56

initial MRI. We see cerebral ad feed that's moderate in

29:59

degree in the right temporal lobe and we do have some reduction

30:02

here of the values. You can see

30:05

here that the lateral infer lateral ventricles are greater than

30:09

99 percentile. We do have some hippocampal volume loss.

30:12

Here's the Ico brain report as

30:15

well. This patient had an amyloid pit

30:18

which was positive. This is amyloid PET CT Fusion. Here's

30:21

the AMOLED pet Mr. Fusion. So this patient had

30:24

Alzheimer's disease next patient was 72 with

30:27

memory loss the presenting MRI here

30:31

and you can see that the values are completely normal

30:34

on neuro plant Quant and I go brain everything

30:37

looks normal for age the patient had an ftg

30:40

brain PET CT. It did show some hypo metabolism

30:43

in the temporal lobes. Although it was not statistically significant.

30:47

But there was statistically significant hypo metabolism in a

30:50

posterior scene that gyride and the parietal lobes when I

30:53

ran it through the mineral analysis

30:56

software. So here's the PET CT here is the pet Mr.

30:59

Fusion and this is just another look here

31:02

that everything's normal on Ico rain instantly the

31:05

patient had a little cavernous angioma here. They came back in 2016

31:08

still not seeing a lot of atrophy, but

31:11

an amyloid pet was done. Here's the PET CT

31:14

Fusion amyloid pet Mr. Fusion and this was

31:17

positive. So this patient has Alzheimer's disease they came

31:20

back in 2018. And now we're starting to

31:23

see attribute. This case is just a reminder that neuro. I'm

31:26

sorry that quantitative volume metrics is really a volumetric

31:29

snap point at one point in time. So it's very

31:32

important to keep ordering serial follow-up want every

31:35

time the patient comes back for neural Imaging. This

31:38

next case was an 81 year old was sudden onset

31:41

memory loss. This is a look at their MRI in

31:44

2020. You see a big infar

31:47

in the right occipital lobe extending into the lingual gyrus

31:50

and the right hippocampus.

31:53

And and so here is the right PCA that's

31:56

occluded. And this is why this patient develops

31:59

sudden onset memory loss. Now the patient actually

32:02

interesting. Hey had a neural Point study

32:05

back in 2014, which was normal. This

32:08

was before they had the infirm and now when they repeated the

32:11

neuro Quant and 2020 after the infrar to

32:14

see that all of these values are statistically significant. So if

32:17

Campbell atrophy and Largemouth at the inferior lateral ventricles,

32:20

this is what the segmentation looks like. You

32:23

can see the infarct here and here you can see the atrophy

32:26

of the right hippocampus due

32:29

to the infarct there so that by the way

32:32

was a patient would obviously that had faster dementia. This next

32:35

patient is a 73 year old male with profound

32:38

visual hallucinations delusions gate difficulties,

32:41

resting Tremor frequent Falls and only minimal memory

32:44

loss. This was their initial MRI.

32:47

It showed some apathy in the occipital lobes

32:50

and in the parietal lobes kind of moderate into

32:53

See this was the neuroquant study in

32:56

2014 and we can see here that we're just

32:59

getting statistical significance and hippocampal atrophy.

33:02

Again, Eiffel brain also shows statistically significant

33:05

hippocampal atrophy and also so some reduction

33:08

in the volume of the occipital cortex and amyloid

33:11

pet study was then ordered. Here's the

33:14

annually PET CT Fusion. Here's the AMOLED pit Mr. Fusion

33:17

and this is a positive study. So the diffuse

33:20

finding of the ambivid Tracer throughout the cortex Talbot

33:23

was also done and it

33:26

showed some tow deposition in the occipital lobes as

33:29

well as in the posterior singulate gyrus

33:32

and in the parietal lobes and MRI

33:35

was repeated. We've got some progression here an atrophy

33:38

of the bilateral parietal lobes and bilateral occipital lobes

33:41

this patient incidentally also had a vestibular schwannoma

33:44

on Imaging and fdg brainpit

33:47

CT was done and it shows cortical hypo

33:50

metabolism. That was statistically

33:53

Significant in the occipital lobes and the parietal lobes also in

33:56

the posterior cingulate gyrus. This is the again another

33:59

look at the fdg protocol hypo metabolism

34:02

in the occipital lobes parietal lobes and posterior cingulate

34:05

gyrus. There's a color map overview. This is

34:08

this pet cfdg pet CT surface

34:11

map where you see hypo metabolism in the parietal lobes

34:14

and the occipital lobes. And so this patient had

34:17

dementia with Louie bodies. These are abnormal deposits

34:20

of alpha synuclean. These are the Louie bodies here

34:23

on pathology and it's less common than Alzheimer's disease.

34:26

So 1 million patients are living with DLB in

34:29

the US versus 6 million with Alzheimer's survival rate

34:32

is typically five to eight years after diagnosis visual hallucinations

34:36

occur in 80% So when you get this as part of

34:39

your clinical history, please be thinking of dld because it's really

34:42

often the first presenting symptom and they can have

34:45

beta amyloid plastic deposition and neural family Tangles and

34:48

DLB as we saw in this case. It was actually enough to turn

34:51

the amyloid positive.

34:53

And to turn the tile pit positive. So another thing that

34:56

want is very frequently used for is multiple sclerosis. This

34:59

is a diagnosis, although obviously the

35:02

patient presents for clinical exam, but the diagnosis is

35:05

often made on MRI. There's often a very characteristic appearance

35:08

on MRI with a Dawson's fingers Etc. We

35:11

encourage quantitative monumentric Imaging for all

35:14

Imaging for Ms. Patients. And in fact

35:17

our refers really love it again, it eliminates that

35:20

report bias and gives the actual volume of plaques in

35:23

the brain which can be tracked over time if there's clinical

35:26

ambiguity sometimes CSF analysis for all go all

35:29

the global plans is perform. But this

35:32

is really what quantitative volumetrics can

35:35

do. It will color code the plaques based

35:38

on their location Lupo cortical pair of

35:41

ventricular infinitorial and deep white matter. This is

35:44

a neural quants, Ms. Example here

35:48

where it will color code increasing facts

35:51

and the sort of orang.

35:53

Red color decreasing plaques and blue and so

35:56

basically this is extremely helpful because it's not

35:59

so hard to do visually when a patient only has

36:02

a few plants, but when there's a moderate or severe plaque burden,

36:05

it is very hard for our eye to see what's new what's enlarging

36:08

what's getting smaller. But you know computers are

36:11

better than humans at pattern recognition. So the Quant actually

36:14

does a great job of this. Here's a

36:17

look at icometrics here here the

36:20

the pre-existing plaques and green plaques are all sort

36:23

of stable in appearance. But that enlarging plaque

36:26

is yellow and you can see how it stands out as we

36:29

scroll through the scene and the red flags are

36:32

the new flags and you can see how the software really just makes this

36:35

jump out.

36:37

This is cortex a look at what a cortex products and

36:40

again on the city. We see the new and enlarging

36:43

packs and reddish orange and the shrinking

36:46

packs in blue.

36:47

This is no look at plaque segmentation with

36:50

neuro Quant. This is what the reports look

36:53

like for multiple sclerosis. This this lesion summary

36:56

here is part of the neuroquant MS report where you

36:59

get the lesion count according to location more importantly

37:02

the lesion burden by volume. So, you

37:05

know as plaques in large, you know, they become confluent

37:08

so the count is not always the most reliable but the the

37:11

burden here is what's very important. So the actual

37:14

volume of plaque so and

37:17

then here is again looking at the lesion Dynamics. It

37:20

gives you the count of new plaques by location the account

37:23

and volume of enlarging shrinking or

37:26

stable plaques and also the T1 hypo intensities

37:29

and then you get a map here of the prior versus the

37:32

current with respect to lesion burden

37:35

and one of the really nice things here is you get a lesion

37:38

Dynamic summary. This can be copy and

37:41

pasted into your report on power scribe or whatever you use or it

37:44

can be pre-populated into Power scribe and it gives us

37:47

summary

37:48

Of what has happened? So 96 lesions with

37:51

the lesion burden of 10.22 and it will

37:54

compare one is in large since the prior study

37:57

and the volume of the enlargement. This is

38:00

that icobrain report for Ms. Where

38:03

again, you get the overall volume of plaque the volume

38:06

change new enlarging shrinking and graphs here

38:09

that demonstrate that this is their pre-populating reporting template

38:12

same thing. This can be pre-populated right

38:15

into your power scribe report which can actually help save

38:18

time. Now, it's very important to know that

38:21

quantitative volumetric Imaging is not the

38:24

same today as it was even like five years ago or

38:27

even two years ago. So these Quant companies

38:30

are really they're primary focus is improving the

38:33

accuracy of segmentation and one of the ways

38:36

that they've done that is sort of shifting from a machine

38:39

learning based approach to a deep learning based approach

38:42

and so it turns out for multiple sclerosis, you know machine learning does

38:45

a great job and filling of lar.

38:47

Replaced like in the period ventricular white matter and

38:50

the Deep white matter, but it's sometimes struggles

38:53

with the inferential plaques and address the

38:56

cortical plaques in part because the input sequence is a flare and

38:59

sometimes for example in for tentorial plaques are better seen on

39:02

T2. So after deep learning was employed

39:05

here at an ensemble method you can see here on the

39:08

original machine learning a couple of these plaques were missed

39:11

here and here but these were then picked up after deep

39:14

learning was a plot then they

39:17

also went on as versions progress. This

39:21

is sort of an ensemble plus attention you net approach

39:24

where it's sort of a third look after a machine learning and

39:27

deep learning are applied. And now we picked up another plaque

39:30

here and these two plaques here. So again, these products are

39:33

just getting better and better over time. Here's cortex.

39:36

This was with machine learning and now with

39:39

deep learning product that's in development. We've now

39:42

picked up these additional two class in the ponds and

39:45

picked up this here. And so

39:47

Getting deep learning I think is going to be really sort

39:50

of the product of the future and probably

39:53

in combination with machine learning for Ms. Patients. And

39:56

in other applications want now.

39:59

You know, one of the concerns that Radiologists have is

40:02

you know, what value does this bring and

40:05

will it cost me time? So it turns out

40:08

actually that it improves detection of disease activity. So

40:11

when you don't use quantitative software in general

40:14

Radiologists will you know

40:17

in this particular study found that 24% of

40:20

class for active in 76 were stable but when

40:23

you use the quantitative post-processing, you

40:26

know, 76% of patients were

40:29

assessed to have disease activity when it was you. So these are the

40:32

same patient population. This is a big difference 24 to 76

40:35

percent. So what was very helpful here, it

40:38

also improves the reliability of reading so

40:41

22% Improvement in inter reader

40:44

variability of Legion count with software 23% Improvement

40:47

and Inter reader variability with lesion count

40:50

software.

40:51

And then you know, it actually can enhance

40:54

productivity. So in this small trial hundred MRIs

40:57

with patients with MS from 11, imaging centers

41:00

50% were read without font and 50% with

41:03

and turns out the neural radiologist was

41:06

able to read eight studies an hour without want and 13

41:09

an hour with font and that pre-populated reporting

41:12

template. So that's 38% more reports per

41:15

hour with font volumetrics in the pre-populated reporting

41:18

template, which is encouraging to know

41:21

that want does not actually cost us time

41:24

in several clinical scenarios. This

41:27

is also a product and development by Michael

41:30

metrics and can actually measure the volume. This is

41:33

just a routine grain MRI this area

41:36

of the upper cervical court is typically included on the

41:39

sagittal sequence. So thank calculate the volume and that

41:42

can help differentiate Ms. Phenotypes and correlate to

41:45

clinical outcomes. Now, we

41:48

also use Quant for epilepsy. I like this

41:51

case this

41:51

The neuroquent this is I go brain. But in this

41:54

case here, you see here that the hippocampal volume on the

41:57

left was within normal limits the hippocampal volume

42:00

on the right was within normal limits, but you see here that

42:03

the asymmetry index was abnormal. It's outside

42:06

of two standard deviations of the mean and this

42:09

can help you potentially, you know

42:12

assess for meteor temporal sclerosis in the subtle

42:15

cases. I mean, we can all see it in the obvious cases but a

42:18

lot of times these are temporal sclerosis can be quite subtle.

42:21

So this is a very useful report. And

42:24

again here is the echo brain report. We also

42:27

can use plant for traumatic brain injury. This is

42:30

a micro brain here calculating the detecting the

42:33

subdural hematoma and calculating the volume of that as well

42:36

as the degree of midline shift and cisternal compression.

42:39

This is their trauma report again

42:42

that highlights those features neural font.

42:46

You can use the triage brain atrophy report to look for

42:49

areas of traumatic brain.

42:51

Series you can also this is micro

42:54

brain and you can see the diffusion tensor Imaging which is

42:57

sometimes applied when a patient has a few sex onal injury.

43:00

So this is a 62 year old professional football player

43:03

with symptoms of chronic traumatic encephalopathy patient

43:06

had pretty severe memory loss. But when

43:09

I looked at his brain back here in 2019, I really

43:12

didn't see any ATV just look completely normal in terms

43:15

of volume, although incidentally. I noted that he had multiple

43:18

sclerosis and he wasn't aware of that diagnosis before so

43:21

quantitative volumetrics was applied. This

43:24

is micro brain, which is negative. This is neuro Quant

43:27

which was negative. Totally normal brain volumes. The patient

43:30

then went on to have an ftg brainpet CT here is the PET

43:33

CT Fusion. Here's the surface map. And here's the pit Mr. Fusion

43:36

completely normal the patient then was

43:39

still having significant memory loss. So they decided to pursue with

43:42

an amyloid pet. Here's the PET CT Fusion. Here's

43:45

the pit MRI fusion and to my surprise the amyloid

43:48

pet was positive. So you see diffuse finding of the Tracer throughout

43:51

The cortex so in the clinical setting

43:54

of CTE trauma has been suggested to demonstrate

43:57

increase amyloid beta peptide levels, although the

44:00

extent of AMOLED deposition and CTE has not been thoroughly

44:03

characterized.

44:05

But if you look here this was an unfortunate in

44:08

18 year old who had fallen and died 10 hours.

44:11

I'm sorry and who died 10 hours after

44:14

the fall the autopsy revealed beta amyloid plaques

44:17

and AMOLED precursor protein.

44:20

So here is the data amyloid plaque here

44:23

on pathology. This is the amyloid precursor

44:26

protein this amyloid planks were

44:29

found within hours after the traumatic brain injury. So there's

44:32

epidemiological association between TBI development

44:35

and Alzheimer's disease. Here's by the way is the APT

44:38

tracking along the damage axons. So rapid beta

44:41

amyloid flag formation may result from accumulation of

44:44

amyloid precursor protein and damage axons

44:47

and disruption of that beta amyloid Genesis

44:50

in cantabulism following traumatic brain

44:53

injury. We also can use plant for

44:56

oncology This Is On Cue neuro, which I had

44:59

alluded to earlier. This is what an On-Q neural report looks

45:02

like in a patient with a GBM. So here's a T1

45:05

Contrast yours T1 plus contrast the patient had

45:08

had the gdm tumoral area

45:11

resected here. But we've got nodular enhancing tissue along

45:14

the resection cavity. We also have nodular enhancing tissue

45:17

in a remote area here. This is what the flare

45:20

sequence looks like. So this is a

45:23

sort of a Time Point progression. Here's at Baseline

45:26

six weeks after the patient received chemical radiation

45:29

here was a resection cavity to remember there was

45:32

some enhancement around that that's the Flair the ADC.

45:35

This is a restriction Spectrum Imaging which

45:38

is an advanced diffusion technique that helps differentiate two,

45:41

true tumor progression from treatment related

45:44

change and then we have Dynamic susceptibility contrast

45:47

here. And so here we

45:50

see the RSI was negative and the perfusion

45:53

there was no perfusion abnormality here.

45:57

This actually was biopsy and you know, this was you know,

46:00

there was no recurrence here then the patient again started developing

46:03

some nodular.

46:05

Tumor here a lot of increase in flare hyperintensity,

46:08

but the RSI remain negative and

46:11

again the perfusion study remained negative. So this

46:14

was pseudo progression three months after initiation of

46:17

immunotherapy and I'm sorry actually this was when the

46:20

patient was biopsy because they just weren't sure if this might be tumor

46:23

progression and it turns out it wasn't it was just pseudo progression

46:26

the patient then came back and they have the satellite area

46:29

of enhancing tissue again, a lot of flair hiker

46:32

intensity, and now all of a sudden the RSI is

46:35

a very positive and we now

46:38

have this perfusion abnormality. So this was true

46:41

progression six months after Baseline and this

46:44

is part here of the quantitative volumetric product

46:47

that RSI now this on

46:50

funeral product here is showing a segmentation where we

46:53

have the enhancing tissue segmented here in green then a

46:56

product for in blue and the surrounding Flair hyperintensity. This

46:59

is the Color Fusion with RSI. Here's

47:02

positivity on RSI. That's the T1.

47:05

Contrast here's a look at the report the tumor

47:08

core which is the enhancing tissue and the necrotic tissue here

47:11

and we're giving the volume of that that carry

47:14

tumoral flare hyperintensity and the

47:17

enhancing tissue alone where you

47:20

get the volume.

47:21

So here's a look here of this patient with this

47:24

tumor with overlaying segmentation.

47:28

And then this was post-treatment. This is what it looked like and overlying segmentation.

47:31

By the way. The blue is the resection cavity. The red

47:34

is the enhancing tissue and the yellow is surrounding zone

47:37

of flare hyperintensity. And you see here on this.

47:40

It's actually an updated on qural report. But that

47:43

enhancing tissues in red we've got the volume of the whole lesion,

47:46

which is enhancing tissue the necrotic four and the surrounding

47:49

zone of flare hyperintensity and the blue is the

47:52

recession cavity. The resection cavity is not included in

47:55

the whole lesion volume. And then we have progression over

47:58

time. The enhancing tissue is gone down post

48:01

property postoperatively as has the whole

48:04

lesion volume.

48:05

This is another example here of a

48:08

patient with a GBM. This is

48:11

before Quant was applied and this is after segmentation.

48:15

And this is sort of a look at

48:18

kind of what you would see if you see me through your images on

48:21

packs. Here's the overline segmentation right

48:24

here.

48:27

This is another neuro-oncology case pre-op and

48:30

post-op segmentation.

48:33

So we can also use quantum for spine

48:36

analysis in particular. It's used for degenerative spine

48:39

disease. This is a Columbo an FDA

48:42

approved product where it's segmenting the volume

48:45

of the vertebral bodies as well as the disc bulges

48:48

and just protrusions here. You see the

48:51

facet arthropathy segmentation and lateral recess

48:54

narrowing. This is another look at Columbo's product

48:57

here where it's measuring the disc Extrusion and this

49:00

is all automated and and then

49:03

a report is generalized is generated

49:06

from the quantitative data here. It's measuring the

49:09

degree of canals stenosis Etc. It will

49:12

also detect fractures and things like that. And basically it

49:15

is generating a report for you. Now, you can modify that

49:18

report so you can sort of upload it into your

49:21

power scribe and then modify it as needed. But all your values

49:24

will be there on canal stenosis at each level. So this

49:27

is very exciting. This is synapticas product,

49:30

which also does want for spine again measuring

49:33

The degree of spinal canal stenosis, they

49:36

debris of naturalist thesis. It's detecting Charles nodes and

49:39

it's also detecting transitional Anatomy

49:42

here, which is very useful and again pre-populating the

49:45

report I can tell you I spent about 70% of my

49:48

day reading degenerative spine reports as a

49:51

degenerative spine studies in having software that could actually

49:54

be the study for me. I would find extremely helpful.

49:57

Here's a look here at a segmentation

50:00

and quantification of the degree of

50:03

neural phrominal narrowing. This is a look here

50:06

on a CT setting where you can segment here and detect

50:09

fractures with the segmentation software. Here's

50:12

Siemens health and ears AI ragged companion

50:15

detecting a degree of Andrew holistic thesis and virtual

50:18

body Heights Etc. So that

50:21

is very useful. And then we can also use AI after

50:24

image acquisition for natural language processing utilization.

50:27

So for example red AI has a product and

50:30

it on it does automated generation of the

50:33

report.

50:33

Question it AI Solutions learn

50:36

and adapt your own report style. So it turns

50:39

out every single radiologist in the world prefers their own

50:42

report style to their colleagues. It's just how we are how we're

50:45

made up as humans. The great thing about this AI tool is

50:48

it it learns and memorizes your style and then applies

50:51

that so you dictate your you know,

50:54

your your findings and Ill it will generate

50:57

your impression for you. It decreases the words. You

51:00

have to dictate by 35% It prioritizes the

51:03

answer to the clinical questions. Sometimes we're moving

51:06

fast. We may forget to answer the clinical question. This software

51:09

will remind us of that. It will also detect any potential Left

51:12

Right errors and highlight those so it's saved times

51:15

it saves time and it can lower Radiologists burnout.

51:19

So the question is is only AI tools we talked about.

51:22

You know, is it what's hype and what's reality vendors have

51:25

high value claims for their AI Solutions.

51:28

But how do we know they deliver? What is Promised? So

51:31

I really believe in the importance of

51:34

multi-center multi-reader blinded clinical validation trials

51:37

and I sort of have devoted a lot of time to that

51:40

as have my colleagues. So this

51:43

is a this is a an abstract

51:46

manuscript in preparation from a trial we did deep learning

51:49

generated syntheticure is interchangeable with

51:52

and better in quality than conventional stir.

51:55

This is some of the images that I showed you earlier. This is

51:58

again, the standards stir and this is a synthetic stress.

52:01

So this is a blinded trial and what

52:04

we found is DL generated synthetic stress fine and our images

52:07

we're interchangeable with the conventionally acquired stir while

52:10

providing significantly higher image quality suggesting the

52:13

potential for routine clinical practice utility. This

52:16

is another trial we did again a manuscript

52:19

and prepped icon base synthetic or inter.

52:22

Visual with and better quality than conventional stir and evaluation

52:25

of trauma. This is a trauma subset here. You

52:28

can see the fracture. This is a conventional stir and this was

52:31

the DL synthetic stir

52:34

here on the right and what we found here is that DL

52:37

generated synthetic stress fine, Mr. Images were interchangeable

52:40

with the conventional acquired stir while providing

52:43

significantly higher image quality and patients presenting in

52:46

that subset with trauma suggesting the potential

52:49

routine clinical practice utility. Here's

52:52

another clinical validation trial we did this

52:55

one is entitled deep learning vendor agnostic

52:58

dicom base reconstruction enables 40% faster spine

53:01

MRI scans which match or exceed

53:04

quality of standard of care and this is a prospective multi-center multi-reader

53:07

trial. This was actually published

53:10

in clinical neural radiology. And what

53:13

we found here was that fast do statistically Superior

53:16

to standard of care for signal to noise ratio and

53:19

imaging artifacts in the spine. So here's a

53:22

Dirt image here's a fast image. And here's the fast enhanced with

53:25

deep learning. And so really our conclusions for

53:28

that DL Master exceeded perceived image quality of

53:31

standard of care spine, Mr. Exams

53:34

enabling 40% scan time reduction, ideal qualitatively

53:37

outperformed standard of care in reduction of

53:40

artifacts and perceived signal to Noise We also

53:43

apply that quantitative ssim which it

53:46

tests to the image integrity and preservation after DL.

53:49

So we knew we weren't introducing any errors or corrupting the

53:52

data with the Deep learning solution and this

53:55

study suggests the potential utility for routine

53:58

use of deep learning reconstruction and

54:01

clinical practice. Now one question that I

54:04

had was is it possible to effectively apply more than one AI solution

54:07

to obtain the benefits of both tools while

54:10

still maintaining image quality and accuracy as

54:13

our paper here deep learning enables 60% accelerated

54:16

volumetric MRI while preserving quantitative

54:19

performance. This is a prospective multi-center multi.

54:22

Trial and this was published in ajnr

54:25

found that afastial was statistically

54:28

Superior to standard of care for perceived quality across every

54:31

single Imaging feature that we looked at despite the

54:34

significant stand-time reduction.

54:36

So deep learning reconstruction allows 60% stand

54:39

time reduction while maintaining that high volumetric quantification

54:42

accuracy consistent clinical classification and

54:45

perceived Superior image quality, when compared

54:48

to the standard of care. So this trial really supported the reliability

54:51

the efficiency and the utility of dl-based enhancement

54:54

for quantitative Imaging and so the hope

54:57

is that these shorter scan times made boost the utilization of

55:00

volumetric quantitative MRI and routine clinical

55:03

setting.

55:04

So, you know just in summary AI Solutions are

55:07

really playing a critically important role in neuroimaging.

55:10

So it's important to understand what tools exist

55:13

out there in the marketplace and how you

55:16

can really tailor these tools to sort of best enhance patient

55:19

care in your clinical practice. Thank you

55:22

so much for joining us. I really appreciate you taking

55:25

the time.

55:26

Thank you so much for that incredible lecture Dr. Bash at this

55:29

time. We'll open the floor for any questions from our audience. You may

55:32

see submit a question for Dr. Bash to the Q&A feature.

55:36

Thank you so much Olivia. Thank you for having me. So I'm going through

55:39

some of the questions here on the Q&A and it looks

55:42

like the first question is what is the future of AI and

55:45

Radiology?

55:46

Well, I personally think that AI is here to stay and

55:49

actually not just in Radiology, but in all occupational sectors,

55:52

so really any areas of Our Lives that

55:55

are impacted by technology. I believe will soon

55:58

be completely transformed by AI within Radiology.

56:01

I expect you know, exponential growth

56:04

in AI Solutions and overall

56:07

Suites. I expect more acceptance

56:10

and implementation of these AI Solutions

56:13

over time and I think that we are going to get to

56:16

a point in the sometime in the near future where

56:19

computers will have a first look at every single

56:22

Imaging exam before a human Lay's eyes on it. Another question

56:26

we have here is Will AI take

56:29

the place of Radiologists. I don't believe so, I

56:32

think that they will work hand in hand in a synergistic

56:35

manner AI tools computers

56:38

are better at pattern recognition than humans are

56:41

but humans are much better at Global reasoning than

56:44

computers. So

56:46

Really the best outcome for our patients and

56:49

diagnostic accuracy. And every other component is a

56:52

synergistic relationship now having said that

56:55

I do see a role for AI in the future for automated

56:58

readings of sort of to separate the normal versus

57:01

abnormal exams, which would apply initially to

57:04

things like chess x-rays just sorting out

57:07

the pile of normals from abnormals and maybe flagging the

57:10

areas of abnormalities. I think we'll probably start in simple

57:13

indications such as Endo tracheal to placement

57:16

things like that. I do see a

57:19

future where potentially AI is

57:22

applied to every single mammogram. I know at radnet.

57:25

We now use AI for all of our mammograms again, it's making

57:28

a huge difference, you know one to you

57:31

know, detecting cancers one to two years earlier than expert Radiologists

57:34

again, because computers have that Advantage for

57:37

pattern recognition. And then I also see in the future someday

57:40

potentially just automated Imaging automated interpretation

57:43

of degenerative spines, you know.

57:46

About 70% of my day reading degenerative spines. It would

57:49

be nice to have a tool that could really help with that so that you

57:52

know, we can spend our time looking at more critical findings and

57:55

things like that. So that's where

57:58

I see that another question I

58:01

have here is which AI tools do you think will

58:04

become standard of care in the future? Well, that's an

58:07

interesting question. I actually think that all of the AI solutions

58:10

that I mentioned in this presentation will at some point

58:13

and in some form become standard of care in

58:16

the future now, I think that what happened in different stages I

58:19

think deep learning for image reconstruction will probably

58:22

be one of the first tools to become standard of care because there's

58:25

just no downside. I mean the patients get in

58:28

and out of the scanner, you know, 60 to 70% faster That's What

58:31

patients rate as the most important thing of

58:34

their Imaging experience is how quick they can get in and out.

58:37

It's better for the Radiologists improve diagnostic accuracy

58:40

because you have higher image quality. So potentially detecting those

58:43

early Mets or you know faster and

58:46

Refer the Imaging Enterprise because now you you know,

58:49

your workflow efficiency is increased you could potentially scan 50% more

58:52

patients a day. So DLR, I think will become standard of

58:55

a care very in a very near future triage tools.

58:58

I think will definitely become standard of care. We

59:01

saw lots of examples how you know, you can save patients

59:04

brain tissue by getting them

59:07

on the table for mechanical thrombectomy earlier for stroke

59:10

patients reduce costs get them in and out

59:13

of the door faster. So these kind of tools I think will be very useful

59:16

and will definitely become standard of care cancer screening

59:20

again. Definitely I think will become standard of

59:23

care in the future at a little bit later stage. I

59:26

think they'll be a widespread utilization of AI and

59:29

front office applications billing Etc. You know,

59:32

those kind of things can be prone to human error and why not

59:35

use computers that you know do a very good job of that

59:38

head angulation. I think will become standard of care want Imaging

59:41

at some point in the future. I would like to see that become standard

59:44

of care and

59:46

also synthetic applications are something that I think we'll see in the

59:49

future just to save time and improve image quality.

59:53

Next question is what is the biggest challenge to acceptance of

59:56

AI Solutions in clinical practice? I think

59:59

you could look at that from either the Imaging Enterprise

60:02

Administration perspective or from the radiologist from

60:05

the perspective of the Imaging Enterprise. I

60:08

would say one of the big questions is which AI

60:11

Solutions best fit your clinical practice whether

60:14

you're in patient or outpatient what provides the

60:17

best value what's real? And what's hype and then

60:20

a big thing is the ROI, you know,

60:23

as you return on your investment, whatever you have to pay to get

60:26

this AI solution is it worth it? And how will it

60:29

impact workflow efficiency? That's a huge issue. Does it

60:32

take time how much time will it take to train the text and the radiologist and

60:35

how easy is implementation? And so

60:38

I can tell you cloud-based AI Solutions are always going

60:41

to be more accepted probably than hardware

60:44

and see, you know, something that requires Hardware installation

60:47

at each site for a place like radnet where

60:50

we have 350 Imaging sites. We don't really want to be

60:53

Only Hardware at every single site now from the radiologist perspective.

60:56

I think the two main challenges are either related

60:59

to a time and trust and time

61:02

is you know, most Radiologists all operate on an rvu basis,

61:05

even if an AI solution offers value, they

61:08

may not be interested in using it if it costs them

61:11

time. So it's very important for AI vendors to remember

61:14

how important it is to not impact workflow efficiency

61:17

either for the for the Imaging

61:20

Enterprise or for the Radiologists that Radiologists

61:23

might also be concerned about does it take time to learn about the products

61:26

and is there seamless workflow integration and as

61:29

customer support available if there's a question about the product and then

61:33

in terms of trust Radiologists, I think have a challenges

61:36

rebate regarding whether they believe in the accuracy

61:39

and reliability and consistency of the

61:42

AI solution and really the value of the product.

61:45

So and next question

61:48

here is you would mention that use deep

61:51

learning and reconstruction throughout your

61:52

Enterprise was implementation seamless. And

61:55

how did you get the Radiologists on board with

61:58

accepting this new AI technology or any other AI tool?

62:01

So what we did with deep learning

62:04

at radnet our CTO Dr. Lawrence Tannenbaum,

62:07

he really was in charge with all

62:10

of the in charge of all the implementation he set up a pilot

62:13

projects on both east and west coast where the section

62:16

heads looked at a real clinical cases the

62:19

standard of care side by side with a fast except,

62:22

you know, an accelerated sequence that had deep learning

62:25

applied and we would just sort of randomly choose a

62:28

different sequence on each patient. So we didn't cost much time

62:31

to the patient at all. They look at both of them and the section

62:34

has it turned out uniformly agreed that the fast DL

62:37

images were Superior in image quality

62:40

and signals noise ratio contest noise ratio, when compared

62:43

to that standard of care, so they're acceptance

62:46

of the product again. It's much better to have people see it

62:49

for themselves than to tell people what to do. So they're acceptance.

62:53

Products through Dr. Tannenbaum's pilot studies really that enthusiasm then

62:56

spread through their whole section and it ended up being just

62:59

really a company-wide acceptance that this was a

63:02

positive change for our company to use deep learning and so

63:05

really, you know integration was

63:08

fairly seamless, you know in terms of acceptance throughout

63:11

the company.

63:13

Next question we have here is someone has asked

63:16

where is post-processing done on the scanner monitor

63:19

at the workstation or possibly on Pax and then

63:22

there's you know, do you need the raw data or can use

63:25

the dicom ETC? Well, I guess

63:28

the this depends on which product you're talking about. For example

63:31

for subtle medical deep learning post-processing

63:34

is performed automatically after every sequence

63:37

acquisition and then it's

63:40

transmitted through the vendors Cloud portable

63:43

portal Friday from that icon data

63:46

and return within one to two minutes. So it's actually returned before

63:49

the next sequence is generally you know, while the next

63:52

sequence is just getting going and so it doesn't require any

63:55

Tech input after implementation has been established and

63:58

it doesn't require any local site Hardware which

64:01

is a huge benefit and that post-processing is

64:04

done all before the images ever arrived to packs

64:07

now for Quant the dicom series are

64:10

input is pushed to

64:12

Vendors portal by the technologist and

64:15

they will designate at that time, you know, what clinical

64:18

indication that they're using, you know for Dementia

64:21

or seizures or Ms. And then that report

64:24

is auto-generated by the vendor typically within

64:27

five to seven minutes. And so by the

64:30

time it even arrives on your packs, the reporters are already there.

64:33

None of us are usually a head on our work list by five

64:36

to seven minutes. So it's there and ready to go for you in packs

64:39

by the time you open the study. So it's really a very seamless component.

64:44

Let's see here. If we have we are five minutes after the

64:47

hour. So if you have to go thank you so much for joining us.

64:50

If you can stay I'll see if there are any other

64:53

questions here.

64:55

This is this question is from a neuro radiologist

64:58

who has used some AI

65:01

companies for neurodegenerative disease presumably

65:04

for Quant and they

65:07

it's a sort of a long question here, but they

65:10

noticed some discrepancy between segmentation and visual

65:13

analysis of easier temporal atrophy.

65:16

What I can say to that is I personally have

65:19

not noticed a big discrepancy of anything. I think

65:22

one of the strong points of quantitative post-processing is

65:25

that ability to reduce bias. There's

65:28

just so much report bias, you know, if I may

65:31

be a moderate person someone at my colleague might be calling it mild another

65:34

calling might be calling it severe I think that there is

65:37

a lot of differences in style when it comes to reporting cerebral

65:40

atrophy. So I think that the Quant is

65:43

actually does a very good job in helping to eliminate that

65:46

by comparing it to a very large normative database so

65:49

that you know, whether the volume of any particular structure such

65:52

as the hippocampus is more than two standard deviations below

65:55

the mean so I actually my

65:58

personal experiences. It's been very helpful in eliminating that

66:01

visual bias that we have next

66:04

question.

66:07

Any scope of career in AI Radiology for Radiologists

66:10

if I'm understanding the that question

66:13

correctly. I think that this needs to be a much bigger issue.

66:16

I would like to see all of our new residents being

66:19

trained in AI, you

66:22

know, just getting familiar with it's what's out there

66:25

in the marketplace. I also think that Radiologists will play

66:28

a very key role in Vendor Development A Products. It's

66:31

it's just so important for these vendors to

66:34

have the input of Radiologists so that

66:37

they know what may be important in our clinical practice. So I

66:40

strongly encourage Radiologists, you know vendors to

66:43

incorporate Radiologists on their medical Advisory board so

66:46

that they can get a feedback from people that are out there in

66:49

the world, you know interpreting these studies.

66:53

Um, let's see here how a radiologist can keep upper hand

66:56

when AI implemented in their institution future of

66:59

Radiologists. I'm not

67:02

quite sure. I completely understand that question. But

67:05

in terms of if you're asking, you know

67:08

how to Radiologists stay up to date on this I think, you know

67:11

attending webinars such as these and keep attending

67:14

the meetings the the big neuro Radiology meetings AI

67:17

is always a very hot topic lots of

67:20

amazing lectures that you'll see at these meetings

67:23

sort of stay abreast of what's going on. It's a

67:26

constantly changing field new and exciting products are coming

67:29

out all the time and not only that but it's

67:32

interesting and it's useful very very useful in our

67:35

clinical practice.

67:37

last question here

67:39

Do you for see a future where an

67:42

organization like ACR can certify or a

67:45

credit AI algorithms just because of the

67:48

proliferation of AI algorithms. That's very

67:51

interesting question. I'm not sure what the role of

67:54

ACR will have in this.

67:57

I think ACR tends to always be a backboard reflection

68:00

of what's happening in the field of radiology. And you

68:03

know, I know that we recently

68:07

did a webinar that you know by sponsored

68:10

by ACR that talks about for example, those pilot

68:13

products that we the pilot that we did the

68:16

pilots that we did at radnet for incorporation of

68:19

deep learning reconstruction, but in terms

68:22

of actual accreditation, I don't know that's kind of a gray Zone

68:26

in terms of that. One of the big things will

68:29

be CMS reimbursement for AI Solutions, you

68:32

know this AI sort of led the way in terms of showing CMS

68:35

that it actually made a difference in morbidity and

68:38

mortality.

68:39

Prompted that new technology add-on payment

68:42

I think as if we can convince CMS

68:45

of the value of a lot of these products, you know,

68:48

hopefully that will help with the ROI in terms of increasing the

68:51

implementation of AI Solutions

68:54

in our practice. So I think that's all we

68:57

have for the questions right now. I wanted to thank you so much for joining

69:00

us and thanks to those people who hung an extra nine or

69:03

10 minutes after the presentation to get the questions answered.

69:06

Have a great day.

69:08

Dr. Bash, thank you again for your lecture today and thanks to all

69:11

for your participation in our new conference.

69:13

A reminder that you can access the recording of today's conference and

69:16

all our other previous new conferences by creating a

69:19

free MRI online account.

69:21

If you'd like to access our full library of case-based topics

69:24

with unlimited see me, you can sign up for a free 7 day trial

69:27

of our premium membership. Be sure to join us next week

69:30

on Thursday, October 13th at 12 pm eastern time

69:33

for a lecture with Dr. Alka Singh Hall on the

69:36

role of Doppler ultrasound and liver cirrhosis and portal hypertension.

69:39

You can register for that lecture at mriline.com and

69:42

follow us on social media for updates and reminders

69:45

on upcoming conferences. Thanks again, and have

69:48

a great day.