This is quite good age detection - other efforts that I've seen (from face images, rather than retina images) are +/- 5 years, whereas this paper gets age to within +/- 3 years or so.
The fact that the model automatically detects anatomical features that humans recognize as being important is promising for a variety of applications, I think.
Any optician can do exactly the same thing during a simple check of your eyes.
In this case I can’t see the benefit of this type of screening. The tool used by the optician is just a cheap lantern, but only the tool needed by the computer is way more expensive.
Perhaps a cheap lens/illuminator attachment for smartphones could be engineered for retinal imaging. If this method could be adapted to phones, many more people could be checked at low cost.
That's like a $40 upcharge at my optometrist (USA). But she is a good family friend, walks me through everything she's seeing on there, and gives me glowing compliments on how beautiful my optical nerve is. It sounds stupid, but I'm going to miss that when I go to my new optometrist.
I don't think anyone would say don't build the AI/ML. Just that the technology already exists. Until then, people should continue to see their eye Dr when one is available to them.
There are a few startups/companies that specialize in going to poor countries and using lens attachments on cell phones to do eye exams. They send in people who aren't doctors, take a picture of the eye, and email/upload the pictures to doctors who can analyze it for issues.
IIRC, in those situations, the problem isn't the accessibility to a doctor, but rather the accessibility to someone who has a device capable of imaging the eye. I imagine this would be the same. Replacing the doctor with an AI doesn't really solve the problem.
It might be way easy to sell to a nineteen year old looking for a gap year activity. "Go to PNG, meet some people, save their life. Look, you just take a photo, and minutes later it says 'They need help'". The doctor reading the image is going to take days or who knows how long.
I'd suspect the tools will eventually be dramatically cheaper than a trained optician. Put it in a pharmacy, grocery store, or other central location and people could screen weekly instead of once every couple years (assuming they even see an optician).
As I have explained before, in Spain an optician does not charge anything for it because it’s ridiculously simple and it’s a small percentage of the effort that an optician needs to do for a prescription.
This test is different of a blood test for example. You don’t need to control the evolution weekly, or even monthly.
This is one of the most simple screening process. Because it’s easy to train the model with tons of images already made by opticians and optometrists. It does not have any value, it’s too simple.
My brother is an optician and has a huge database with retina images. If even me with limited knowledge of ML and my brother as field expert could train a model to detect blood pressure problems.
There are plenty of Americans who don't have access to a "family doctor". Not everyone can afford or bothers to get an annual checkup.
Same reason grocery stores often have a "check your blood pressure" kiosk, or offer flu shots. It's quick, it's convenient, and you reach people who would otherwise miss out. Sure, a doctor or nurse can check your BP or give you the vaccine, but there's a significant benefit to expanded access.
AI will not fix your healthcare system.
Pharmas will charge people as much as they want to access to this screening service. They have been doing this for decades.
Why would pharmaceutical companies be involved in this?
As you say, it's a super easy scan needing minimal equipment - a light and a camera. I can get a free blood pressure reading at my grocery store via an automated kiosk - no nurse or doctor required. This would likely be similar.
>but there's a significant benefit to expanded access.
I agree that expanded access to diagnostic tools is great, but access to diagnostic tools is only part of the problem. Access to care following your diagnosis is just as, if not more, important.
What are you going to do after this kiosk tells you that you have heart disease? The proper answer is: go see your doctor. But if you're someone who is in a situation where you don't have access to a family doctor, your situation has not improved.
> What are you going to do after this kiosk tells you that you have heart disease? The proper answer is: go see your doctor. But if you're someone who is in a situation where you don't have access to a family doctor, your situation has not improved.
First, this scan doesn't give you a diagnosis. It tells you you're at higher risk for a heart attack. There are free lifestyle changes that can be made to reduce that risk once it's known.
Second, lots of low to middle income folks have to decide what to spend money on. An annual routine eye exam might not meet the "must have" threshold. A "you might have a heart attack" might make it worth going to the doctor and paying the copay.
Simple? Yes but there's absolutely value in this. In Pakistan I've never had an opthalmologist comment on my retina before, despite wearing glasses since I was a child.
Not everyone has an opthalmologist brother, or is a machine learning engineer. You should have used this to your advantage to make something like this.
How many startups make simple applications? Uber is a simple marketplace, as is Airbnb, and PayPal is a glorified accounting database. Please don't downplay the impact of a product because of its simplicity.
AI and ML should take mankind to a new level. This article describes something that opticians and optometrist have been doing for a long time with a more classic computing algorithmic approach.
Morbidity and mortality can be reduced with the prescription of beta blockers, aspirin, statins, and ace inhibitors. +\- clopidogrel. Being at risk isn’t fate: you have modifiable risk factors that can be mitigated (eg, weight, cholesterol, blood pressure.)
Additionally, CPR training of your immediate family can be - literally - lifesaving in the event you do have that heart attack.
It gives a high-risk group the ability to notify their friends and family and practice what to do in event of a heart attack. My mom frequently stays with me, but outside of "call 911" I don't really know what to do in teh event of a heart attack. Time lost is lives lost, so better preparation could quite literally save lives.
I suspect that it's mostly the amount of data that allows this, and that the software is just using run-off-the-mill statistics/ML. The title should reflect that. It's not the software, it's the data.
> opening a new opportunity for artificial intelligence in the vast and lucrative global health industry.
well done! google will join the lucrative industry of the people’s health. the rich will survive, the other will build the machine to make them survive.
i think it'll be quite some time until ML and CV can do the breadth of diagnostics for all possible things.
my wife is an optometrist. just from looking at retinas, she finds brain tumors among other scary systemic problems. even if you have perfect vision, go in for an exam, it can literally save your life.
Those are two entirely unrelated methods. One worked, one failed. One is a sort of "wisdom of the crowd" algorithm (as in human crowds) and it failed after initially showing promise (essentially it worked until someone told the crowd), one is machine learning from images.
Honest question: why is that an alternative medicine eg. Iridology is seen as quackery but a computer predicting heart attack risk is not? I know the science behind AI but from a patient point of view, they both analyze the eye.
47 comments
[ 4.1 ms ] story [ 104 ms ] threadThis is quite good age detection - other efforts that I've seen (from face images, rather than retina images) are +/- 5 years, whereas this paper gets age to within +/- 3 years or so.
The fact that the model automatically detects anatomical features that humans recognize as being important is promising for a variety of applications, I think.
In Spain (where I live) most opticians do not charge anything for this test, because it’s a minimal part of a typical prescription.
It’s a very simple test. It’s not a ML or AI challenge at all.
IIRC, in those situations, the problem isn't the accessibility to a doctor, but rather the accessibility to someone who has a device capable of imaging the eye. I imagine this would be the same. Replacing the doctor with an AI doesn't really solve the problem.
So you'd still need the optician/doctor, if only because they have necessary equipment.
There are so many things that can happen to your vision, at high life cost. Get your vision checked regularly.
This test is different of a blood test for example. You don’t need to control the evolution weekly, or even monthly.
This is one of the most simple screening process. Because it’s easy to train the model with tons of images already made by opticians and optometrists. It does not have any value, it’s too simple.
My brother is an optician and has a huge database with retina images. If even me with limited knowledge of ML and my brother as field expert could train a model to detect blood pressure problems.
Too simple, no value.
It's also a first step proof-of-concept towards harder diagnoses.
It’s not an ML or AI challenge at all.
Same reason grocery stores often have a "check your blood pressure" kiosk, or offer flu shots. It's quick, it's convenient, and you reach people who would otherwise miss out. Sure, a doctor or nurse can check your BP or give you the vaccine, but there's a significant benefit to expanded access.
You've moved the goalposts so much that I'm not sure what your point is anymore. Was it just to boast about Spain?
As you say, it's a super easy scan needing minimal equipment - a light and a camera. I can get a free blood pressure reading at my grocery store via an automated kiosk - no nurse or doctor required. This would likely be similar.
I agree that expanded access to diagnostic tools is great, but access to diagnostic tools is only part of the problem. Access to care following your diagnosis is just as, if not more, important.
What are you going to do after this kiosk tells you that you have heart disease? The proper answer is: go see your doctor. But if you're someone who is in a situation where you don't have access to a family doctor, your situation has not improved.
First, this scan doesn't give you a diagnosis. It tells you you're at higher risk for a heart attack. There are free lifestyle changes that can be made to reduce that risk once it's known.
Second, lots of low to middle income folks have to decide what to spend money on. An annual routine eye exam might not meet the "must have" threshold. A "you might have a heart attack" might make it worth going to the doctor and paying the copay.
Not everyone has an opthalmologist brother, or is a machine learning engineer. You should have used this to your advantage to make something like this.
How many startups make simple applications? Uber is a simple marketplace, as is Airbnb, and PayPal is a glorified accounting database. Please don't downplay the impact of a product because of its simplicity.
Additionally, CPR training of your immediate family can be - literally - lifesaving in the event you do have that heart attack.
well done! google will join the lucrative industry of the people’s health. the rich will survive, the other will build the machine to make them survive.
my wife is an optometrist. just from looking at retinas, she finds brain tumors among other scary systemic problems. even if you have perfect vision, go in for an exam, it can literally save your life.
In https://www.kaggle.com/c/diabetic-retinopathy-detection actually ML algo did better than humans. Yet it will take time for apply these widely.
How is that different from just measuring the blood pressure (with really inexpensive device)?