Comment by bastijn | Hacker News Reader

Comment by bastijn | Hacker News Reader

In my line of work we say: "It ain't cancer until the pathologist says so."

Basically it is a joke amongst the pathologists that no matter what MD does the diagnosis, s/he will always send a biopt to the pathologist and ask if it is cancer. Well, in simple terms of course.

I guess the correct moment to say you have been diagnosed with cancer is when the pathologist tells the physician it's cancer.

Beautiful thing is neural networks / deep learning can really outperform the pathologist now that we have introduced digital pathology. Sure they will for some time give the final verdict but before too long AI will do the actual diagnose with better accuracy than a human possibly could in a fraction of the time. Reports started coming in the last year. E.g. [0][1][2]. Last one is directly mentioning years.

0) https://www.extremetech.com/extreme/233746-ai-beats-doctors-...

1) https://news.developer.nvidia.com/diagnose-heart-disease-wit...

2) http://bigthink.com/paul-ratner/heres-when-machines-will-tak...

Good luck ever convincing doctors (or the general public, legislators, etc.) to let themselves be replaced by machines. Working in healthcare, everything ML/AI related has to be worded very carefully as something that helps doctors spend time doing other things, rather than replacing what they do. People aren't super comfortable with the idea of AI in medicine yet, and physicians have a history of being extremely protectionist when it comes to their jobs.

What's going to happen when this starts becoming a reality is that doctors will start bringing up Dr. House level medical sleuthing tales as proof that machines can never replace physicians in medicine. Obviously this will just be confirmation/survivorship bias, but few people even know what those terms mean. I bet it will take decades for AI to be integrated with medicine after it's ready to do so.

We are doing it today. We of course sell it as assisting the doctor. Partly because of things like FDA and legislation, partly because doctors and reluctant to change. In due time they will value the algorithms decision over their own and the balance will shift. It is like the chess computer. They claimed it could never beat them, they tried, and they lost. Our algorithms run against millions of other cases in milliseconds. A number a human brain cannot possibly contain no matter how experienced. And each new case adds to its power. Not only the cases one doctor sees, but that every doctor in the world sees. It's numbers really.

Healthcare becomes the next industrial revolution. Of course nothing is ready yet. Not only doctors but also law needs to change. When an algorithm makes a false negative, who will we sue?

Note I do not say doctors will no longer be needed. They very much will be. But their work shifts towards the social part of the job. Guide the patient, help him/her understand. Assure. Etc.

> People aren't super comfortable with the idea of AI in medicine yet, and physicians have a history of being extremely protectionist when it comes to their jobs.

How much of this is due to the fact that the "AI" we have right now is only competent to speed up tasks for a human? Apple and Google have invested enormous amounts of money in GPS navigation, yet it still can't deal with the sort of random road closures/change in traffic patterns that are common here in D.C. For about two years, Uber's GPS directed drivers to the side of my office building that not only has no entrance, but is on the through-way. Drivers would blindly blow by me and sometimes ended up in Virginia.

To use a different example, circuit layout is something that seems optimally suited to AI. Yet its still a competitive edge in the industry to be able to manually lay out critical parts of your CPU. (Apple does it with the Ax).

To be perfectly honest, I don't see much in the way of people complaining about computers replacing their jobs. The media, for example, is totally unskeptical about the prospect. They focus entirely on how to find new jobs for people. What I do see a lot is people playing up quite primitive technology as if it'll replace humans any day now.

I've worked in a medical imaging lab before, there are currently real, working models that are able to e.g. accurately segment organs in 3D to find info such as volume. And of course there are diagnostic tools as well. You're complaining about GPS apps with inaccurate information. That's kind of a non-sequitur.

I mention GPS routing because it's been subject to incredible investment yet are still reliant on humans to deal with routine unexpected events.

I don't see the point of your organ example. That sort of data analysis seems like the sort of things computers do routinely; how does that address the need to have a human in the loop? To use a different example, software for doing analysis of car and plane designs has been the subject of decades of development. Yet engineering teams for new car and plane designs are bigger than ever.

My phone also knows that there is an accident on my route home, messages me before I get into the car, and tells me to take an alternative route. I cant remember the last time my GPS didn't navigate directly to my destination.

Another way to phrase the question: Would you rather use a paper map or GPS?

But both a paper map and a GPS are designed to aid a human operator. This thread is about replacing the human operator.

Actually the automobile and aviation sectors are under more strict regulation than Healthcare because the ECU software and aviation software running, when making a mistake, does not only influence one life but many. So whilst being added more and more in cars and planes, you will see AI faster in Healthcare than in cars or planes. I've been in a shared European subsidiary project recently (http://www.emc2-project.eu/) and these sectors were always laughing when I complained about FDA. I wouldn't want to work a single day in their sectors. They had to disable the second core on a dual core fpga because it was not allowed. Meanwhile, I was happily running my data science in a cloud network.

Regarding design, design is a peculiar thing potentially much more difficult to automate than a medical diagnosis because there is only emotion attached. For medical diagnosis for a lot of cases it is just comparing to others and spotting the differences in the images as early as possible. Before even a human eye can spot them. And not with one dimension but hundreds. Many algorithms do not give the actual diagnosis, they let the doctor know this scan deviates from the default in marked places so a doctor can assess quickly without searching.

> Regarding design, design is a peculiar thing potentially much more difficult to automate than a medical diagnosis because there is only emotion attached.

"Emotion" is not the reason companies still design airplanes and CPUs by hand.

True. Was for the cars. The other part holds for airplanes. By the way, I know of at least one case where a cpu algorithm steered the team on optimizing the resistance of a plane wing (or whatever needs to be optimized there, it was a conversation over beers). Example was that they never could have thought of that improvement themselves. Basically you had to be very stupid to try it as it couldn't work, until a spinoff (one of the millions of automated tries) actually did. CPUs I have no knowledge of.

"Good luck ever convincing doctors (or the general public, legislators, etc.) to let themselves be replaced by machines."

This presumes doctors are in control, which they are not. I suspect the insurance industry will adopt and impose these technologies as an independent analysis channel in the same way they impose current bureaucratic process.

I do agree this will be a slow process.

This is closer to reality than you might think. Doctors get less and less control in this as Healthcare costs continue to rise due to the aging population. Healthcare has to be come more efficient to be able to keep up or nobody would be able to pay for it anymore. So where we sold to doctors Five years back we now sell to hospital MT. Hospital MT controls the money and are constantly searching for ways to perform more cases a day. AI is their only viable option and so they will buy in sooner than later.

Im not sure we could ever replace the Human element of receiving a bad diagnosis with machine.

I'm firmly for AI augmenting and bettering a Drs diagnosis ability.

Respectfully disagree. The moment people are able to reliably say that the computer is more accurate than a human, people are going to want it IMO. Cancer is just too scary to fuck with, they are going to want the fastest, most accurate thing you can get. And then they will want a second opinion from a human, of course. :) But that's no different than today!

Reply on news.ycombinator.com