Hacker News Reader

Comment by rurban | Hacker News Reader

Mass spectromity is a bit expensive. My company can detect a common plant virus reliable with multi-spectral imaging only. Normal hires B/W cameras. 10 days before visible symptoms. SpexAI.com

Only vision scales. We did mass spectromity in Formula 1 because we had the money and it doesn't need to scale.

View on HN

Such approaches are cool for plants and maybe animal healthcare, where the other reasonably priced alternative is nothing.

I don't know how I'd feel about reliable precision tools in human healthcare being substituted with B/W cameras and AI. It reminds me of a certain car company, where the human lives lost aspect wasn't so liked, even if they saved some money.

This is doing correctly what the car company did wrong. They start with a system that does not have a fatal incident upon failure. Meaning that even 1% or 10% failure rate is totally fine and possibly positive in terms of ROI.

On the other hand, even 0.01% is not acceptable when the system is responsible for human lives.

I also believe that cameras and visible spectrum are scalable and sufficient, and there is evidence it is so. The most reliable system driving cars is the homo sapiens and apart from a bit of proprioceptive, vehicles are operated based on vision sensors(eyes).

To conclude, I find it a clever move to start refining the technology in a field where ROI is in dollar terms and not in lives saved.

To the found: Good luck, i am jealous.

>On the other hand, even 0.01% is not acceptable when the system is responsible for human lives.

I've got bad news for you about a whole lot of medical tests and interventions if 0.01% is unacceptable for you. I get your point that the stakes might be higher with medical technology than in some other fields, but you're still setting the bar too high.

The first pathogen I saw mentioned by name while skimming through the linked article was H. Pylori, where the false-positive and false-negative rates for various currently used tests are several dozen times worse than that at about 0.30-0.90% [0]. Even just being hospitalized in and of itself has failure rates. Just eyeballing some of the numbers at [1] and elsewhere, I suspect that at least 0.01% of hospitalizations result in a death attributable to a hospital acquired infection.

[0] https://www.aafp.org/pubs/afp/issues/2019/0701/p16.html [1] https://www.cdc.gov/healthcare-associated-infections/php/dat...

It could be used for disease control and not necessarily personal health care. You could use something like the plant thing if it worked for people by photographing an entire stadium of people and finding infected people, airport entry lines, etc.

> B/W cameras and AI.

You forgot multi-spectral. That means single hires images for a lot of frequency ranges, which is better than a single color image, which merges together all frequencies at once.

In the end the accuracy of detections, false-positives, false-negatives wins. And this not comparable to this certain car company which refuses to use proper sensors. we also do infrared, 3d pointclouds to get the angle at each pixel, and thermal imaging.

another neighbor of us is doing real-time imaging for face detection, which is the industry leader world-wide. They can install a lot of cheap cams in football stadiums and give you a list of all 80.000 people, when they have access to the government passport/driver license photo database. they have. but they cannot detect virus infections as we can do. nor bacterial infections. only if they do something illegal. who threw a bengal fire or started a fight.

This. Mass Spectrometry should portably only be used to confirm/train AI vision models. Once a large enough data set has been created for a particular pathogen, incremental diagnostic cost would be effectively 0.

You’re right, I didn’t notice the multi-spectral part. My point still stands, I wouldn’t want human diagnostic standards to slip but it is exciting technology. Thanks for sharing.

A reason for bw cam is sensitivity. No bayer filter means you capture all the photons. There are cameras that are sensitive enough to count photons as well. When I was in the lab we used EMCCD cameras (electron multiplier ccd) that we used to detect very low signals from cells (we wanted to be nice to them, if you hit them with UV light they don't like that).

There are all kinds of really great hacks to eke out an extra couple % at this level, if you consider you're counting photons. Back thinned, reverse mounting, high bandwidth multichannel readouts. Most of those trade offs are a thing of the past, but in the 2000s we were hacking the beans off those cameras to get them to do stunts for us.

Things are fast good AND cheap now though!

Sounds like this is exploiting a correlation between the virus and how the plant reacts to it. That's very specific.

The nice thing about mass-spec is that you can use it as a dragnet detection method.

>Normal hires B/W cameras. 10 days before visible symptoms.

*visible to the human eye

Reply on news.ycombinator.com