- It's reading from (mostly) the surface, not the bulk mass. Not great for heterogeneous things like pills
- It uses machine learning models on 10-datapoint IR reflectance spectra, meaning it's only useful in a trained regime. It doesn't give information about composition, but instead classifies a sample as a member of a pretty constrained population. So a mystery substance that can't be roughly identified ('vegetable', 'pill' etc) can't be scanned, etc. If nobody's built a model for the thing you're scanning and for the property you want to evaluate, then you're out of luck
- So, evaluating drugs ("we can distinguish fake from real viagra") is done by looking at the surface coating which typically has no active ingredients (and even if there were, the signal would be washed out by inactive ingredients). The model is basically trained on 10-100 scans of a presumed good viagra pill, or maybe 10-100 different good viagra pills if they felt like it
- Building models requires purchase of a $250 license in addition to the $250 hardware, which is just ridiculous. Of course they're doing the calculations on their servers, but it still seems really scammy, hostile to developers, and counterproductive to launching an ecosystem of scanning models. The useless 10-point IR "spectra" notwithstanding, I would totally buy one of these if you could use open data and open models supported by a public community.
I'm not quite interested enough yet to buy bare Hamamatsu (or neospectra/other knockoff?) spectrometer chips, but would totally buy a breakout PCB for one if somebody had a bunch made
They've just finished a round, but they're reasonably regular.
Cheap diffraction grating and a cheap camera; basically, the idea would be to somehow use an IR sensitive camera (most are - at least in the far-mid range), an IR source (maybe an unfocused IR laser diode?), and a diffraction grating.
Ideally for the grating you'd use one for IR - but I haven't found one that wasn't reflective-based (instead of transmissive) - there's probably a good reason (likely having to do with cheap materials and IR absorption - which is why lenses and mirrors for laser cutters tend to be pretty pricey).
Anyhow - thems the basics. Take that, get an image from the camera, run a fourier transform on the data, get the peaks, then pass it thru a trained CNN (?) - heck, you might be able to forgo the transform part and just use the data from the camera directly (repurpose imagenet or something).
Yeah - I think this whole thing could be made an open-source project; probably even an instructable...? At the very least, it could become an interesting science fair project for some enterprising kid...
As that "enterprising kid" who couldn't wait until the SCiO was completed, and who was fed up with his 3d printed visible light spectroscope, I put in a lot of research effort figuring out how to go about creating a portable, financially attainable NIR spectroscope. I started out thinking it would be fairly straightforward since I already worked on creating CCD sensor driver boards in my early highschool years (after all it just seemed like you'd need a linear CCD with no IR-cut filter, a diffraction grating, a prism with 95%+ reflectivity, and an IR emitter with a peak wavelength of ~920nm). I invested about 6 months working through the various roadblocks I encountered until I managed to get half-decent results from it when taking the spectrum of various salts. Still never managed to fully refine it since I didn't have the budget to (Heck the Eagle schematics I sent off to a PCB fabricator were so poorly-designed that in order to get it to work correctly it required me to solder a wire across two traces to prevent the sensor from blowing up). It's still somewhere in one of my drawers and is a fun novelty item (and it did win (second?) place at my junior science fair) but at least with my design, it really wasn't the portable tricorder I was hoping for...
The healthtech company I work for has expressed interest to CP about using the SCiO phone for a project. If that doesn't work out, we might replicate this work by Fraunhofer. We're already looking into it. And since it's not core to our business, if we do, we'll open source it too.
There are two classes of problems in spectroscopy: qualitative (trying to identify what the sample is) and quantitative (deciding concentrations of one or more analyte in sample).
Qualitative is quite simple: each material have quite unique fingerprint and you match it against dataset.
Quantitative is very very tricky. You have to pick good region of spectra and you have to pick good features (say peak areas or peak high). For complex solutions you will get a lot of peak overlaps and then you have problem with matrix effects for most samples (non-linearities caused by combinations of various compounds and so forth).
Good news is that a lot of interesting things can be found in near infrared region, so you don't really need expensive diffraction gratings and delicate optics.
- All spectroscopy is surface reading, with NIR penetrating somewhat deeper. But it's still up to you to prepare the sample.
- IIRC, there were 400 datapoints corresponding to 1 nm spaced wavelengths from 700-1100 nm. The models aren't super smart, correct, and need a lot of collective data to be useful. They were able to do more than you describe though (a demo they did was scan cheese for protein/fat composition, though to be fair those are trained models, probably with external data).
- It's up to you (and the other data providers) to prepare the sample correctly and test it the same way.
- Yeah, I don't know what they are thinking on this.
Don't get me wrong - I didn't find SCiO particularly useful and am not sure where it's going. But I wouldn't be surprised to see technological improvement either. That is, if they have a business model and further funding.
Namely, it can tell your body fat from a scan of your skin, and the results seem pretty accurate (same result as a Withings scale for example).
Now that's interesting. I wonder if they've considered marketing it as a way of testing the purity of illegal drugs?
The important data and algorithms are held in their servers, not in the phone app. This gives them a lot of control and responsibility over the models, leading me to believe that they wouldn't store data on illegal drugs.
I also feel it's kind of scummy to keep this on the server. There is no reason that this functionality couldn't be offline and on the phone, except to let them charge you monthly for the hardware you're effectively leasing from them, rather than letting you use the device you own.
Also, I expect if this does take off that someone will replace their app's phone home behavior and sever-side components with an equivalent on localhost. Slightly later will be the cloned scanners that plug into the app. They should focus on providing higher-quality and higher-accuracy scanners than competition can achieve, rather than trying to lock out competitors.
Everything about this campaign has been dubious. One of the first videos of an "unboxing" was a close friend of the creator who, after deleting his video, went around posting comments about how it was just "good fun" and "innovation takes years", etc. Every demo of the product is a staged example of very low hanging fruit, pardon the pun. The actual utility of this device seems deeply suspect.
That said, it sounds like harmless fun.
But if you really care about accuracy, its not that hard to assess the quality of fruit and vegetables with your own senses, don't need a spectrometer/iphone contraption.
I don't doubt it's possible there could be some correlation, but it seems like a big leap from estimating carbohydrate content to detecting food-borne illness or contamination. A quick google turns up this comment (of uncertain reliability!) saying most food safety issues are caused by toxic animals or plants, pathogenic microorganisms, or chemical contamination.
Somewhat relatedly, inorganic ions don't have nir activity, only their complexes with organic counterions have nir activity. And there are like a gazillion different things that complex with lead/mercury/etc. in unpurified biomaterial, so direct nir measurements of a sample are not a productive approach.
That being said, it would certainly be a big moneymaker if placebo contaminant detectors became a thing with Chinese consumers like surgical masks for smog
I disagree here - as someone on a ketogenic diet, the sugar content of food is extremely important to me. Labelling standards are abysmal, too; it's legal to market something as "0g carbohydrates" if there is less than a 0.5g carbs per serving. Considering that serving sizes can be very small, that can be extremely misleading.
For instance - a bag of pork rinds may be listed as 0g carbs, even going so far as marketing that fact on the front of the bag, while the third ingredient by weight is sugar. If the bag claims that there are 20 servings, then the only real information you have is that there are less than 10g of carbs in the bag. When you're on a diet that limits you to 20g per day, that's a huge problem.
anyhow spectral sensing could be one of the main stream for IoT sensing