Hacker News Reader

Comment by Aeolos | Hacker News Reader

Fancy that, my current project is to create a low-cost, high-quality optical coherency tomograph for private practitioners. I am planning to open-source the software and, depending on how things turn out over the next few months, possibly also the hardware designs.

This device can immediately detect around 80% of the pathologies of the posterior eye segment. It's high time we made this part of first-line diagnostic exams.

View on HN

I don't know if you have an organization around you or not, but be careful. If your software is actually producing some sort of a quantitative or even qualititative "you have this visual defect or not" then the FDA may very well consider you a Medical Device and you'll have to go through submission, etc.

There are a lot of small businesses out there that are unwittingly (and illegally) producing medical devices and don't even know it (yet!).

We will be going through the long and arduous FDA and CE approval process. This is a big part of the cost.

Best of luck with that. As someone who has taken medical imaging software through both, my preference would be to get CE marking first and allow the market and the device to become establish before going for a 510K. The FDA predicate device process together with the glacial speed of the agency means making modifications to a system is unnecessarily difficult which is something you will need to do when innovating in a market. I find the CE process to be a more pragmatic approach.

On some of the darker days with the FDA - and despite their 90 day target we had over 900 days between submission and final clearance - I found it helped to remember the famous Churchill quote; "When you are going through hell, keep going."

Thanks for the advice. We already have experience with the CE process, it's painful but doable. The manufacturing company has been cleared for FDA before, so for now I remain optimistic. (Famous last words...)

That sounds amazing. What's the price point you're targeting?

I think we paid ~$13K for our Atlas OCT (Placido Disk), and ~$43K for our Cirrus OCT (Spectral domain). Both refurbished, a few years old.

I can't go into too much detail (publications pending), but the initial prototype cost roughly $6K in components. The basic idea is find the optimal balance between hardware-vs-software and speed-vs-cost-vs-quality. The device attaches to a regular slit lamp, so that shaves off a big part of the cost too (no need for a separate examination table and optical assembly).

This is spectral-domain OCT with tracking, so quality falls somewhere between Cirrus and Spectralis. However, I'm working on a few sweet mathematical tricks that actually surpass anything available on the market right now - hardware and software has advanced hugely over the past few years.

We already have a hardware manufacturer and are currently fighting to secure the necessary funding. This technology has to make it out there.

Great to hear you're helping to make that happen! So many medical devices (in clinical or research settings) are ridiculously outdated, expensive, and newer tech can often leapfrog the diagnostic capabilities. I've done research, mostly in soft tissue research, and it's crazy how much potential there is out there.

What's your system's A-Scan rate (SD-OCT, so I guess somewhere betweek 10kHz to 50kHz)? Sensitivity? Also I guess you already know, that there's a group that aims at producing a hyper low cost full field TD-OCT (they aim at far below 100USD per unit) intended to be used as a device that's given the patient for self administered diagnostics in ongoing treatment monitoring. There was a talk about this at the recent BiOS 2016 conference.

Personally, I'm working at exactly the other end: Ultra-High-Speed OCT. 1.5MHz and higher A-Scan rate (our group still holds the record of 20MHz equivalent A-Scan rate for a 60°×60° @ 1900x1900 scans posterior segment OCT system).

Awesome, high-speed OCT is amazing. The hardware is still super-expensive, but that is bound to change over the next 5-10 years. Do you have a publication on your system?

Right now I'm working with 70K ascans/s (we've also built a faster swept-source prototype, but the cost-benefit analysis doesn't work out just yet). We have higher-than-average losses due to the slit lamp, so sensitivity is slightly lower than fully-custom systems. We make up for that in post-processing.

I find amazing how far things have progressed in just 10 years. In a decade, we went from struggling to acquire and process x10^2 ascans/s, to real-time processing of x10^5-10^6 ascans/s.

> Do you have a publication on your system?

https://www.osapublishing.org/oe/abstract.cfm?uri=oe-18-14-1...

> to real-time processing of x10^5-10^6 ascans/s.

That would have been me :) (the real-time processing on the GPU part), second author on the paper:

https://www.osapublishing.org/boe/abstract.cfm?uri=boe-5-9-2...

If you want to, you can buy our systems from our university spin-off company: http://optores.com/index.php/products/18-mhz-oct-system

Ah, thanks, I'm familiar with those papers and the spinoff. Great job! :)

I just find it mindblowing how efficient GPUs are at processing this kind of data. I've written implementations for a 6-core DSP (40K ascans/s), modern 4-core CPUs (~70K), 12-core server CPUs (~180K), my laptop's AMD GPU (>120K), a desktop Nvidia 980 GPU (more ascans/s than I care to benchmark.) From a price/performance/simplicity point of view, the GPU outperforms anything else out there right now.

Right now I'm putting this extra computing power to improving quality in other ways than the regular OCT signal processing pipeline. 1M+ ascans/s OCT is great because it opens up new use-cases - I'd love to work on those one day in the future - but there's still a surprising amount of untapped potential in 100K OCT (in the sense that 100K + better processing can actually give better image quality than 1M + typical processing.)

Reply on news.ycombinator.com