Computer vision AI to interpret medical changes

00:00:05:27 - 00:00:08:08

Good morning, everyone,

and thank you for the opportunity

 

00:00:08:08 - 00:00:10:13

to be here in front of you

to talk about this topic.

 

00:00:12:09 - 00:00:15:27

We'll cover broadly these five things

during the entire next 30 minutes,

 

00:00:15:27 - 00:00:19:19

and I'll try to go over

some of the things that the two sessions

 

00:00:19:19 - 00:00:21:02

in the beginning have set context

 

00:00:21:02 - 00:00:25:05

on much around deep learning

as well as a lot around error analyzes.

 

00:00:25:05 - 00:00:25:27

Honestly,

 

00:00:25:27 - 00:00:29:01

I have a sweet tooth and I saw the bear

on Toblerone for the first time

 

00:00:29:08 - 00:00:31:06

So that is interesting.

 

00:00:31:06 - 00:00:34:15

But we'll quickly talk about the image

landscape in health care, the acquisition

 

00:00:34:28 - 00:00:36:03

and some pre-processing.

 

00:00:36:03 - 00:00:36:22

What is the real

 

00:00:36:22 - 00:00:40:09

opportunity landscape of how these things

are being pursued in the industry?

 

00:00:40:21 - 00:00:44:04

Some basics of computer vision,

which I'll probably skip through

 

00:00:44:19 - 00:00:48:15

if time permits i still cover

but I want to spend more time on two

 

00:00:48:15 - 00:00:53:25

use-cases of how this is being used

in the medical for the

 

00:00:55:05 - 00:00:57:02

assessment of the images

 

00:00:57:02 - 00:00:59:05

to take decisions

on the clinical criteria.

 

00:01:01:03 - 00:01:04:21

But before I start a quick overview,

so I'm part of Optum

 

00:01:04:29 - 00:01:07:26

and Optum a part of the UHG,

which is the health services,

 

00:01:08:00 - 00:01:12:02

a health organization

that's a Fortune five company

 

00:01:12:02 - 00:01:15:02

with more than $300 billion revenue

 

00:01:15:02 - 00:01:18:08

in FY 2022

 

00:01:18:08 - 00:01:20:27

and Optum is a leading health services

 

00:01:20:27 - 00:01:23:22

and information and technology enabled

health services player in the market

 

00:01:24:01 - 00:01:26:29

with businesses across Optum

Insight, Optum Health and Optum Rx

 

00:01:27:12 - 00:01:30:22

that leads in transforming

the health care space.

 

00:01:31:13 - 00:01:32:04

Let's talk

 

00:01:32:04 - 00:01:33:23

about clinical images first.

 

00:01:33:23 - 00:01:38:24

So medical imaging normally creates

images of various parts of the body

 

00:01:39:03 - 00:01:42:28

to be able to help with detection

of certain kind of image conditions.

 

00:01:43:09 - 00:01:45:28

There are a variety

of these clinical images that come in

 

00:01:46:03 - 00:01:48:21

and just to relate to a few examples,

ultrasounds

 

00:01:48:28 - 00:01:52:14

really are used, for like one example,

for detecting kidney stones.

 

00:01:54:12 - 00:01:57:05

Radiology X-rays,

I think just to relate to something

 

00:01:57:05 - 00:02:01:04

very recent is all about

detecting pneumonia and COVID

 

00:02:01:15 - 00:02:02:18

and with

 

00:02:03:03 - 00:02:05:16

MRI is helping us more understanding

 

00:02:05:18 - 00:02:08:25

the symmetry of the bone

and also other bone related conditions.

 

00:02:10:13 - 00:02:12:27

But beyond images

that come from radiology,

 

00:02:12:27 - 00:02:16:26

there are also other kinds of images,

like photographic images, where certain

 

00:02:16:26 - 00:02:21:15

clinical conditions can be identified

through visual assessment, for example,

 

00:02:21:15 - 00:02:22:24

eyelid drooping or certain

 

00:02:22:24 - 00:02:26:25

cosmetic procedures that require a certain

kind of medical necessity checks.

 

00:02:27:04 - 00:02:31:24

but irrespective of the kind of image,

it still requires deep clinical expertise

 

00:02:32:01 - 00:02:36:03

to be able to correctly interpret

the image, detect the disease,

 

00:02:36:20 - 00:02:39:08

understand the progression of disease

in applicable cases,

 

00:02:39:18 - 00:02:42:24

but also compute certain

objective criteria to understand

 

00:02:42:24 - 00:02:45:02

maybe the severity

or the extent of the disease.

 

00:02:45:16 - 00:02:46:19

And that requires us.

 

00:02:46:19 - 00:02:49:13

And that requires

great technical expertise. Now.

 

00:02:50:16 - 00:02:52:10

Unfortunately,

 

00:02:52:11 - 00:02:55:10

one thing that happens

is that this is subjective decision

 

00:02:55:10 - 00:02:58:12

making that vary

significantly from one clinician to another,

 

00:02:58:26 - 00:03:01:09

based on their experience,

based on their field of practice.

 

00:03:02:19 - 00:03:06:00

And what this leads to

is a variation in the delivery of

 

00:03:06:00 - 00:03:10:12

health care to intended users

or intended members, and their

 

00:03:10:12 - 00:03:14:10

impacting health outcomes.

The use of AI can significantly augment

 

00:03:14:10 - 00:03:18:23

human decision making over here, driving

high consistency in the decision making,

 

00:03:18:23 - 00:03:22:20

making it more accurate,

reducing that variation,

 

00:03:22:26 - 00:03:26:26

and also ensuring this gets done

more optimally at a lower cost.

 

00:03:27:13 - 00:03:31:27

To give an example of a recent study,

the Babylon AI solution

 

00:03:32:21 - 00:03:37:01

had an 80% overall accuracy on processing

these on generating intelligence

 

00:03:37:01 - 00:03:41:00

from these images with up to 98% accuracy

on certain diagnoses

 

00:03:41:06 - 00:03:44:26

which are very common in primary care

conditions, primary care setting, sorry.

 

00:03:45:09 - 00:03:47:29

And when that same study

or same set of images were analyzed

 

00:03:47:29 - 00:03:51:19

by a set of doctors who participated

in that study, their accuracy

 

00:03:51:19 - 00:03:56:24

varied from 64% to 94%,

which is again just an example of really

 

00:03:57:01 - 00:03:59:08

talking about the kind of variation

that exists

 

00:03:59:08 - 00:04:03:01

and leads to inconsistent delivery

and then impacting the health outcomes.

 

00:04:05:17 - 00:04:06:26

But one of the most important part

 

00:04:06:26 - 00:04:10:02

on all of this is image acquisition

Now for a health care company,

 

00:04:10:02 - 00:04:12:01

there's an abundance

of these images available

 

00:04:12:01 - 00:04:14:02

that come through into the enterprise,

 

00:04:14:02 - 00:04:18:03

through a lot of processes across

the entire health care lifecycle - right,

 

00:04:18:03 - 00:04:21:27

from seeking authorization,

whether a service can be rendered

 

00:04:22:05 - 00:04:23:12

to being able to reimburse

 

00:04:23:12 - 00:04:27:02

a provider for renting that service site,

all of these images come through.

 

00:04:27:13 - 00:04:30:10

But in other scenarios, they can also

be sourced from external images.

 

00:04:30:29 - 00:04:35:14

However, what's still more important is

not just good volume but also good quality,

 

00:04:35:23 - 00:04:39:11

which is most of the time or in many times

not enough.

 

00:04:39:11 - 00:04:42:03

That requires

pre-processing. Pre-processing

 

00:04:42:03 - 00:04:44:26

in augmentation really helps us

improve the quality of images,

 

00:04:45:06 - 00:04:48:27

but then also create variance or,

let's say, synthetic images

 

00:04:49:06 - 00:04:50:08

that allow us to create more

 

00:04:50:08 - 00:04:53:10

generalized models that can perform

better or larger sets in future.

 

00:04:54:05 - 00:04:56:16

Now, while good volume of images

 

00:04:56:16 - 00:05:01:06

is there while quality can be improved,

what is still is a dearth of

 

00:05:01:06 - 00:05:05:05

is labeled annotated data

because this requires clinical expertise,

 

00:05:05:05 - 00:05:06:22

as I talked about in the previous slide,

 

00:05:06:22 - 00:05:09:24

not just interpretation,

but even to train these algorithms

 

00:05:10:00 - 00:05:12:21

that clinical expertise is extremely important.

 

00:05:12:21 - 00:05:13:20

And that is where

 

00:05:14:28 - 00:05:15:15

if not done

 

00:05:15:15 - 00:05:18:17

properly, it impacts the outcomes of what

your model performance is going to be.

 

00:05:18:29 - 00:05:22:06

Just explain this on a very simple view

What you see on

 

00:05:22:06 - 00:05:25:05

the slide is a sample

of an annotated image of maxillary sinus.

 

00:05:25:09 - 00:05:28:06

The frontal area,

uh, these light blue color

 

00:05:28:06 - 00:05:31:03

is the bone area, and the black part

you see is the air cavity.

 

00:05:31:14 - 00:05:35:15

When we want to detect sinusitis,

essentially what we're trying to detect is

 

00:05:35:15 - 00:05:38:16

what is the percentage of air cavity

that is blocked by mucus.

 

00:05:38:26 - 00:05:43:17

Now, imagine if I'm unable to correctly annotate

 

00:05:43:20 - 00:05:44:28

the area of interest,

 

00:05:44:28 - 00:05:48:11

then what can it do to my predictions

later during my assessment

 

00:05:48:29 - 00:05:51:14

There are some techniques mentioned

previously in agumentation

 

00:05:51:22 - 00:05:53:11

I do not want to do

as they have been covered

 

00:05:53:11 - 00:05:56:11

to some extent in the previous

two slides, previous two sessions actually

 

00:05:57:18 - 00:05:59:10

. Now going beyond

 

00:06:00:06 - 00:06:02:27

annotation, let's look at some broad

use cases in the industry,

 

00:06:03:08 - 00:06:05:14

at least in the health care industry,

where these are being used today.

 

00:06:05:25 - 00:06:07:04

One prior authorization.

 

00:06:07:04 - 00:06:09:16

So in the US health care system, essentially it's

 

00:06:09:16 - 00:06:12:05

such a tightly controlled,

regulated system by the government

 

00:06:12:05 - 00:06:14:20

having a set of standards

about how the entire process works

 

00:06:15:11 - 00:06:17:24

Now prior authorization is where

 

00:06:18:13 - 00:06:21:26

for certain specific procedures

or high costing procedures

 

00:06:23:02 - 00:06:26:11

in short it requires providers

 

00:06:26:11 - 00:06:29:27

to seek a prior auth or a pre-approval

before their service can be rendered.

 

00:06:30:07 - 00:06:33:23

Now, what this requires is assessment

at the insurers,

 

00:06:33:23 - 00:06:37:11

end to be able to ensure medical necessity

is there to render the service.

 

00:06:37:11 - 00:06:40:09

And this can lead to material improvement

in the health outcome

 

00:06:40:22 - 00:06:41:28

and that is the prior process.

 

00:06:41:28 - 00:06:44:01

So advanced computer vision systems

 

00:06:44:28 - 00:06:48:13

based clinical decision support system

to extract objective criteria

 

00:06:48:13 - 00:06:52:15

from the radiographic or photographic

images has been meaningfully helpful

 

00:06:52:15 - 00:06:54:27

to assist the reviewers

in the decision making.

 

00:06:54:27 - 00:06:58:05

But sometimes in these medical records

where these images are sourced from,

 

00:06:58:08 - 00:07:01:03

they are also additional context

that is situated in the text

 

00:07:01:12 - 00:07:04:17

so extraction of both the image

objective criteria and text provides

 

00:07:04:17 - 00:07:08:08

with them a more comprehensive

understanding or assistance of what

 

00:07:10:03 - 00:07:12:26

can help them to make right their decision.

 

00:07:12:26 - 00:07:17:03

Another example I take is on claim

processing. In the dental claim process,

 

00:07:17:03 - 00:07:21:27

again, when the claim comes

in, it is supported with dental X-rays

 

00:07:22:03 - 00:07:24:05

and there are classes of dental X-rays

right from

 

00:07:24:05 - 00:07:27:00

like whether it's from the left side,

right side and all that stuff.

 

00:07:27:27 - 00:07:30:26

But when a claim has to be processed,

two assessments have to be made?

 

00:07:31:14 - 00:07:35:11

First one is what is the guideline

of the image that is required

 

00:07:35:19 - 00:07:36:25

for that specific case?

 

00:07:36:25 - 00:07:40:10

And is the dental X-ray that has been provided

compliant with that guideline

 

00:07:40:19 - 00:07:43:05

and necessary to process that claim?

 

00:07:43:05 - 00:07:45:00

So that has been a very,

 

00:07:45:17 - 00:07:45:27

I would say,

 

00:07:45:27 - 00:07:49:24

prominent use case in the industry,

not just dental, but even in other areas.

 

00:07:49:24 - 00:07:53:21

Third is form processing computer

vision is being used today along with NLP

 

00:07:53:21 - 00:07:57:20

to extract entities from claim form images.

You know claims are lengthy forms

 

00:07:58:04 - 00:08:00:17

where a lot of information

has to be extracted .

 

00:08:00:17 - 00:08:01:20

Whether these are extracted

 

00:08:01:20 - 00:08:06:01

for just data entry purposes to digitize

and make it a workable object

 

00:08:06:01 - 00:08:10:15

all to extract certain things to fit into

an automation is use case specific.

 

00:08:10:21 - 00:08:14:10

But one example over here is identifying

which checkboxes

 

00:08:14:10 - 00:08:17:29

have been selected so that you know

what have the provider has

 

00:08:19:00 - 00:08:21:01

basically filled in and you can extract

 

00:08:21:01 - 00:08:24:00

those specific features or inputs

and figure the further processes.

 

00:08:25:04 - 00:08:28:12

Context extraction - even beyond images,

 

00:08:29:11 - 00:08:32:20

medical records

that come as an image, right

 

00:08:32:20 - 00:08:36:26

intelligence medical document processing is decomposing

that unstructured document to extract

 

00:08:36:26 - 00:08:40:19

context of what that description

 

00:08:40:19 - 00:08:42:24

or the medical statement

is trying to convey.

 

00:08:42:24 - 00:08:45:20

And that is further

helping the reviewers in their decision

 

00:08:45:20 - 00:08:47:26

making, along with the visual

input that they are getting.

 

00:08:51:07 - 00:08:53:27

Now I will not covered too much over here

but again

 

00:08:54:07 - 00:08:58:02

image at a very simple

level is a set of matrices and numbers.

 

00:08:58:11 - 00:09:01:01

But those numbers really have to mean

something before

 

00:09:01:01 - 00:09:05:17

they can be put to good use

and for your AI purposes.

 

00:09:06:00 - 00:09:08:29

They are various techniques out there

which help you create those features.

 

00:09:09:09 - 00:09:10:19

Histogram of oriented

 

00:09:10:19 - 00:09:14:10

gradients is one of the most frequently

used features that perform really well

 

00:09:14:22 - 00:09:17:11

because it not only helps

you understand the difference

 

00:09:17:11 - 00:09:20:25

between pixels, but also the orientation

of those, the gradients of those.

 

00:09:20:25 - 00:09:25:20

And that has meaningful implications,

I would say, for X criteria as well.

 

00:09:26:18 - 00:09:29:27

Various types of networks

like convolutional neural networks and

 

00:09:30:24 - 00:09:31:21

vision transformers

 

00:09:31:21 - 00:09:35:16

today, they do a pretty good job

in terms of assessing all those features

 

00:09:35:16 - 00:09:37:10

and really predicting the outcome class

 

00:09:37:10 - 00:09:41:09

quite effectively. Both have a difference

in their underlying base architectures,

 

00:09:41:09 - 00:09:46:16

where CNN's commonly used filters

to be able to convolve

 

00:09:46:24 - 00:09:51:00

the entire image in patches

and create features through those filters.

 

00:09:51:00 - 00:09:53:23

Whereas vision transformers

normally split an image into a patch

 

00:09:54:06 - 00:09:57:25

into multiple patches and basically

flatten them and feed them into the

 

00:09:58:11 - 00:10:01:11

desired input dimension to enable

 

00:10:01:11 - 00:10:02:21

the prediction of right kind of classes.

 

00:10:02:21 - 00:10:03:13

But both are good

 

00:10:03:13 - 00:10:07:05

examples of transfer learning

where pre-built learning is being used.

 

00:10:07:16 - 00:10:10:09

Pre-built learning on a certain

kind of outcome is

 

00:10:10:09 - 00:10:13:20

being used to really learn

or train a model for a similar class.

 

00:10:15:28 - 00:10:16:22

And why I wanted to

 

00:10:16:22 - 00:10:19:14

really skim through those two slides

because I think those are basics.

 

00:10:19:14 - 00:10:21:09

But I also got covered

in the first two sessions.

 

00:10:21:09 - 00:10:22:16

But what really is important

 

00:10:22:16 - 00:10:26:21

to understand, two use cases of how

this has been used in the industry today.

 

00:10:26:24 - 00:10:30:12

Let's look at the first one,

which is the spondylitis problem.

 

00:10:30:12 - 00:10:32:24

Now, what is the problem

in the first place?

 

00:10:33:15 - 00:10:37:25

This is a situation where a certain part

of a bone in your spine

 

00:10:37:28 - 00:10:41:06

called a vertebra, moves forward

in simple terms,

 

00:10:41:08 - 00:10:44:17

a sort of slip disk.

 

00:10:44:17 - 00:10:49:00

The requirement over here for us

was how to identify

 

00:10:49:00 - 00:10:53:01

which vertebra slip has happened

and what is a displacement of that

 

00:10:53:13 - 00:10:57:13

and can we create an objective measure

to really express that displacement?

 

00:10:58:01 - 00:11:02:28

Now, such images come in

abundance in a single document

 

00:11:03:24 - 00:11:08:07

because the MRI will decompose the

I would say the spine structure

 

00:11:08:07 - 00:11:10:12

from multiple angles

and present a lot of images.

 

00:11:10:12 - 00:11:13:12

But what is the most relevant

image over here is the

 

00:11:14:13 - 00:11:17:01

midsagittal view of the lumbar region.

 

00:11:17:01 - 00:11:20:23

Essentially, think of it this way,

if I take a plane and just slice myself half

 

00:11:20:23 - 00:11:23:28

the spine view that gets created

is the view I'm talking about,

 

00:11:24:08 - 00:11:27:09

that is the most relevant,

most appropriate view of being able

 

00:11:27:09 - 00:11:29:06

to make this assessment

in the first place.

 

00:11:29:06 - 00:11:32:26

Now, when you get a lot of these

images of the spine, the first idea,

 

00:11:32:29 - 00:11:38:10

the first step for us really is to use

object detection models to identify

 

00:11:38:10 - 00:11:41:21

which is the right image

that can be used for further steps.

 

00:11:42:05 - 00:11:44:25

But once you've identified that,

I think we've used YOLO framework

 

00:11:45:00 - 00:11:48:28

to be able to create object

detection models for the first part.

 

00:11:49:15 - 00:11:53:00

Now, in this segment,

the second step is for us that we have

 

00:11:53:08 - 00:11:58:09

prepared a segmentation model to identify

where these individual vertebra are.

 

00:11:58:09 - 00:12:02:01

So first to annotate the images

which we have created

 

00:12:02:01 - 00:12:05:15

mask for each of the vertebra,

and then fed into this segmentation model

 

00:12:06:01 - 00:12:09:19

based on the unit architecture

with mainstreamed on imagenet

 

00:12:10:12 - 00:12:12:15

pre trained weights , imagenet weights to get to this.

 

00:12:12:15 - 00:12:16:23

Now over here, I think what to call out

is that I've highlighted two specific

 

00:12:16:23 - 00:12:20:16

vertebra in green and red,

but the model will basically try

 

00:12:20:16 - 00:12:22:12

to predict all of the vertebras

 

00:12:22:12 - 00:12:25:15

why just two over here, one to make it

less confusing for the conversation.

 

00:12:25:15 - 00:12:26:08

But secondly.

 

00:12:27:16 - 00:12:28:09

This green

 

00:12:28:09 - 00:12:30:21

vertebra, which is the fifth

one in the lumbar region, is called L5

 

00:12:31:16 - 00:12:33:25

and the first one is called the sacrum.

 

00:12:35:00 - 00:12:37:26

Most of the listhesis that happens,

happens at this point.

 

00:12:38:12 - 00:12:40:12

So I'm trying to explain this concept from here.

 

00:12:40:28 - 00:12:45:22

Once we have identified

what is the vertebra of interest.

 

00:12:46:06 - 00:12:49:05

The next step is to understand

how much is the displacement.

 

00:12:49:29 - 00:12:52:09

So far, my predictions

have happened to an extent over here.

 

00:12:52:19 - 00:12:55:28

Now I'm getting into the second part

of my use case, which is

 

00:12:56:11 - 00:12:59:11

what is the displacement

or what is the objective criteria?

 

00:12:59:14 - 00:13:05:08

So just to explain this, think of it,

I'll draw one tangent across my lumbar L5.

 

00:13:05:08 - 00:13:10:01

A second, tangent to the sacrum S1

that we have looked into.

 

00:13:10:01 - 00:13:12:16

And then a horizontal plane

across the lower vertebra

 

00:13:13:14 - 00:13:16:25

the intersection of these tangents

from both the tangents

 

00:13:16:25 - 00:13:19:18

on the horizontal plane

creates a certain distance

 

00:13:20:07 - 00:13:22:26

that the ratio of that distance

 

00:13:22:26 - 00:13:25:07

to the overall width of the lower vertebra

 

00:13:25:17 - 00:13:29:06

is what a clinician defines for us

that is the objective measure

 

00:13:29:15 - 00:13:33:20

that could be of material

use for them and this displacement is then

 

00:13:36:27 - 00:13:38:02

reviewed by those

 

00:13:38:02 - 00:13:41:08

clinicians during a prior-auth process to be able to say,

 

00:13:42:06 - 00:13:44:08

'Is this medically necessary?'

 

00:13:44:08 - 00:13:47:20

Is this displaced enough to be able

to approve a certain kind of treatment?

 

00:13:48:01 - 00:13:50:21

And that is how utilization management

normally happens in the industry

 

00:13:50:28 - 00:13:52:19

in this specific case.

 

00:13:52:19 - 00:13:55:23

So now, again,

certain basic metrics over here.

 

00:13:56:02 - 00:13:59:03

One thing which is most important

for us was really being the accuracy

 

00:13:59:03 - 00:14:02:01

by which we are able

to locate the vertebra.

 

00:14:02:12 - 00:14:04:08

So you define dice coefficient.

 

00:14:04:08 - 00:14:08:10

Essentially, it's an reference

score in a sense where of all the pixels

 

00:14:08:10 - 00:14:11:24

that are of interest

that were in the ground root for us, that

 

00:14:12:02 - 00:14:14:14

we took image for L5 and S1,

 

00:14:15:05 - 00:14:17:19

we have looked at what are the

 

00:14:17:19 - 00:14:20:24

What is the area of overlap

on the predicted pixels?

 

00:14:20:24 - 00:14:22:17

And we are trying to basically

just compute

 

00:14:22:17 - 00:14:26:28

that as a base and efforts two times

the area of overlap versus

 

00:14:26:28 - 00:14:29:15

the total number of pixels in the ground

truth and duplicate elements.

 

00:14:29:15 - 00:14:31:03

Essentially in this one score,

 

00:14:31:03 - 00:14:34:00

pixel accuracy is not really

a good measure in such scenario.

 

00:14:34:00 - 00:14:36:22

Again, class imbalance

and that is nothing different

 

00:14:36:22 - 00:14:39:02

than what we do

in, I would say, structured data.

 

00:14:39:14 - 00:14:43:01

But you just go to

those numbers at 0.82

 

00:14:43:20 - 00:14:44:28

does that for reference purposes.

 

00:14:44:28 - 00:14:47:14

But the Dice coefficient here

was a more suitable measure for us

 

00:14:47:14 - 00:14:51:29

to really get a sense of accuracy

that these solutions are basically creating.

 

00:14:51:29 - 00:14:54:13

And in general other use cases in the industry.

 

00:14:58:25 - 00:15:01:03

The other example I want to talk about.

 

00:15:01:03 - 00:15:04:00

So the previous one was more

of a segmentation solution based solution,

 

00:15:04:01 - 00:15:07:11

and this one is more regression

based solution, orthognathic surgeries.

 

00:15:07:24 - 00:15:09:27

So what this is really.

 

00:15:09:27 - 00:15:12:00

Orthognathic surgery,

essentially is jaws surgery

 

00:15:12:11 - 00:15:16:10

that looks at the deformities in the jaw

and their teeth between jaw

 

00:15:16:10 - 00:15:20:04

and their teeth and helps align them

so that they function the way they should.

 

00:15:20:18 - 00:15:23:12

And these surgeries

sometimes also lead to changes

 

00:15:23:12 - 00:15:26:08

in the facial structure of your appearances.

 

00:15:27:11 - 00:15:29:20

There is a specific type of,

 

00:15:29:20 - 00:15:32:06

I would say, X-ray images or radiology images

 

00:15:32:06 - 00:15:36:02

that are materially useful for this kind

 

00:15:36:02 - 00:15:39:04

of assessment called cephalometric X-rays

 

00:15:39:16 - 00:15:42:00

and they are basically

 

00:15:43:04 - 00:15:43:24

very proficient

 

00:15:43:24 - 00:15:48:12

in being able to do a good assessment

of x-ray of craniofacial region.

 

00:15:49:15 - 00:15:50:07

Similar to the

 

00:15:50:07 - 00:15:53:12

previous situation

where when we have this kind of condition

 

00:15:53:12 - 00:15:56:15

documents coming in

or we see a lot of these images.

 

00:15:58:25 - 00:16:00:16

The first objective is to identify

 

00:16:00:16 - 00:16:04:09

the right lateral supplementric x-ray,

which can be used for these purposes.

 

00:16:04:22 - 00:16:05:27

So again, object detection

 

00:16:05:27 - 00:16:09:12

models help you identify those images

with certain position.

 

00:16:09:27 - 00:16:13:13

Once you have identified the images,

what is really is

 

00:16:13:13 - 00:16:16:28

a supplementric assessment over here

is that there are certain points,

 

00:16:17:19 - 00:16:22:01

I would say, on the skull area

and the soft tissue area.

 

00:16:22:15 - 00:16:24:22

And those points are.

 

00:16:24:23 - 00:16:26:18

That is how a doctor really would

 

00:16:26:18 - 00:16:28:28

look at these points set up manually

 

00:16:29:20 - 00:16:32:08

and you see the small green

dots on the screen.

 

00:16:32:08 - 00:16:34:24

Yeah, these are certain points

 

00:16:34:24 - 00:16:37:08

which are on the area

where what you look at is

 

00:16:37:08 - 00:16:40:13

a doctor would create these points

and then try to create a few angles,

 

00:16:41:05 - 00:16:45:17

sagittal analysis and vertical analysis

that really are defined clinical standards

 

00:16:45:26 - 00:16:51:15

to understand whether this situation has

developed and requires the surgery.

 

00:16:52:13 - 00:16:52:23

Couple of things

 

00:16:52:23 - 00:16:56:16

to note over here

is that this is a very time taking

 

00:16:56:16 - 00:16:59:27

task on a doctor from their perspective

 

00:16:59:27 - 00:17:02:13

and still manually cumbersome. So.

 

00:17:03:22 - 00:17:06:19

Solutions over here

are able to help you plot these

 

00:17:07:25 - 00:17:10:07

predict this basically

 

00:17:10:07 - 00:17:12:23

points on the on the image

 

00:17:13:02 - 00:17:17:08

and then we draw these angles as per

the guidance which again are not

 

00:17:17:08 - 00:17:21:06

being used to really take a decision

or establish a condition.

 

00:17:21:06 - 00:17:24:25

Because one thing to note

is that when you talk of faces over here

 

00:17:25:12 - 00:17:28:00

and these kind of angles,

they may have different kind of,

 

00:17:28:19 - 00:17:29:29

I would say.

 

00:17:31:12 - 00:17:34:12

A different outcome

by age, by gender, by ethnicity.

 

00:17:34:22 - 00:17:38:19

So you cannot generalize angles over here

to really explain a certain condition

 

00:17:38:19 - 00:17:39:19

has happened or not.

 

00:17:39:19 - 00:17:42:02

So we then create this kind of an assistance

 

00:17:42:15 - 00:17:45:02

offer this to a clinician

who would normally look at these

 

00:17:45:10 - 00:17:50:05

and then take their own decision

of whether this, uh, surgery

 

00:17:50:05 - 00:17:54:11

should be approved, should

we act on this in the first place does it materially solve

 

00:17:54:11 - 00:17:57:28

or improve the health outcome

that is being desired for this patient.

 

00:17:58:23 - 00:18:03:08

Now two or three important things

I would talk about in addition over here.

 

00:18:05:13 - 00:18:07:25

The way I look at it is that.

 

00:18:08:26 - 00:18:11:00

First of all, these are all augmentations.

 

00:18:11:14 - 00:18:14:13

AI has gotten really good at doing

 

00:18:14:21 - 00:18:17:25

routine task that human

can't do, there's no doubt about that.

 

00:18:18:12 - 00:18:22:13

But I think tasks that impact human life

can't be left completely to machines,

 

00:18:22:22 - 00:18:26:29

and that is where being able to augment

human decisioning with the help of

 

00:18:26:29 - 00:18:30:24

AI can be a significant boost,

both in terms of reducing variability.

 

00:18:31:15 - 00:18:34:12

I highlighted a study or experiment

in the beginning,

 

00:18:35:20 - 00:18:39:12

both in terms of being more accurate

in improving health outcomes,

 

00:18:39:22 - 00:18:41:27

but also doing it

more optimally at a lower cost.

 

00:18:41:28 - 00:18:45:02

Think of this with the US health care

system

 

00:18:45:12 - 00:18:47:21

produces billions of claims every year.

 

00:18:48:22 - 00:18:50:24

And that's only medical claims.

 

00:18:50:24 - 00:18:53:02

There could be other kinds of pharmacy

claims and stuff like that.

 

00:18:53:02 - 00:18:55:15

And this number is in billions

for the US health care system

 

00:18:55:27 - 00:19:01:00

and there is significant radiologist

burnout, significant pressures on doctors,

 

00:19:01:00 - 00:19:05:12

providers, clinicians to be able to manage

this administrative workload.

 

00:19:05:25 - 00:19:10:10

Industry has been focusing a lot

on being able to use technology

 

00:19:10:22 - 00:19:13:14

to enable these clinicians enable

the providers

 

00:19:14:00 - 00:19:17:00

to get rid of administrative burden,

but focus their time

 

00:19:17:00 - 00:19:18:21

more on treating the patients.

 

00:19:18:21 - 00:19:21:10

And that is where these assistants

are being widely adopted

 

00:19:22:09 - 00:19:23:13

to, widely

 

00:19:23:13 - 00:19:26:05

adopted as human augmentation

or decision support tools

 

00:19:26:15 - 00:19:30:04

to really help

the US health care practitioners.

 

00:19:30:21 - 00:19:33:07

The other part also is that.

 

00:19:33:16 - 00:19:36:24

We always are very because this impacts

human life.

 

00:19:36:24 - 00:19:40:25

And I would say from

even a general perspective,

 

00:19:41:15 - 00:19:43:23

we are big advocates of responsible

use of AI

 

00:19:43:26 - 00:19:46:03

So even while selecting

all of these samples.

 

00:19:47:06 - 00:19:49:23

I think it's important

for any practitioner

 

00:19:50:06 - 00:19:53:15

to ensure that there is a good representation across

 

00:19:53:15 - 00:19:57:26

the various kind of age

demographic ethnicities

 

00:19:58:07 - 00:20:01:29

so that the models are not biased towards

one kind of population.

 

00:20:02:09 - 00:20:05:24

They work equally well for a Hispanic

or an Asian or an American and so on.

 

00:20:06:07 - 00:20:10:11

Otherwise, what will happen is there

will be disparity in delivery of care,

 

00:20:10:23 - 00:20:14:10

which again, was another basic

principles of ensuring good health

 

00:20:14:10 - 00:20:15:00

care for everyone.

 

00:20:15:00 - 00:20:16:15

And that is what the government

really wants to do.

 

00:20:16:15 - 00:20:18:25

And I think every health care

provider wants to do

 

00:20:18:25 - 00:20:20:18

take good care of their members.

 

00:20:20:18 - 00:20:23:04

So I think those are the two or

three most important things

 

00:20:24:06 - 00:20:26:14

which are necessary over here.

 

00:20:28:13 - 00:20:30:18

Other than that

 

00:20:30:27 - 00:20:33:12

Yeah, I think this is pretty much

what I wanted to cover,

 

00:20:33:13 - 00:20:35:21

but it's more about the right

kind of use cases

 

00:20:35:21 - 00:20:37:16

which can drive these outcomes again.

 

00:20:37:16 - 00:20:42:05

One last point,

these technologies are not there

 

00:20:42:05 - 00:20:45:09

to replace doctors per say

because I'll just quote

 

00:20:45:09 - 00:20:47:27

my previous statement 'decisions

that impact human life

 

00:20:48:10 - 00:20:52:19

can't be left to machines,

but they can significantly accelerate

 

00:20:53:06 - 00:20:57:07

how AI can be adopted across

much broader industries and also manage

 

00:20:57:07 - 00:21:00:04

the administrative workload

in the health care industry.'

 

00:21:02:01 - 00:21:02:18

Yeah.

 

00:21:02:19 - 00:21:03:02

I think I'll.

 

00:21:03:02 - 00:21:05:28

I'll take a pass, but anything I can.

 

00:21:07:12 - 00:21:09:13

Address. Yep.