Google Health licenses its AI breast cancer screening tool to a medical biz

Google has licensed its AI breast cancer screening model to a commercial medical technology company, paving the way for the system developed by researchers to be tested in real clinical settings for the first time. The partnership with iCAD, a maker of cancer detection and radiation therapy equipment headquartered in New …

[ZaphodHarkonnen]

Good luck to them, they'll need it. Automated cancer detection from mammograms is viciously difficult. Especially as you can't diagnose from an image. The images just help you identify people that will need a biopsy which will be the diagnostic method.

I worked for a company in the field and even with millions of mammograms available for research, detection of cancer was in the 'active research ongoing' bucket.

The two biggest things I would want to know about Google's model are

* Does it use the FOR PROCESSING images or FOR PRESENTATION

* How does it perform on images from BAME ethnic groups

The first question refers to how the mammogram and tomogram images are stored and transmitted. By default manufacturers apply their own 'secret sauce' to adjust the raw image to highlight certain structures for radiologists. Not only do these algorithms differ between manufacturers but they differ between models within a manufacturer. These would be the FOR PRESENTATION images. FOR PROCESSING are generally larger and about as close as you can get to the raw image. They contain the energy detected by the dector with only the most basic of processing done. Stuff like trying to remove the halo effect and other x-ray weirdness.

If the Google models were trained on FOR PRESENTATION images then it will fail as you will have to retrain the models for every single manufacturer and model out there.

The second question is important as so much of research like this is done on people of a white european ethnicity. There are differences in breast composition between ethnic groups. Such as the ratio of dense and fatty breast tissue. This has an impact on how easy it is to identify suspicious structures in mammograms and tomograms. For example increased dense breast tissue not only increases the chance of breast cancer but also masks cancerous structures in mammograms, making it harder to spot possible cancers.

So yeah, I wish them luck. They'll need it. It will be cool if they can make it work.

[Dante Alighieri]

and another thing..

I agree with your points entirely.

And have looked at your posting history at Vulture Central {RIP VC UK}

As a schmuck [sp?] who previously used these images and has enough training to comment, it is a useful adjunct to flag a "but did you see this...", never AutoPilot machine diagnosis.

Given that I am currently attending a major conference virtually, the ethnic mix of lots of stuff is being actively questioned in many imaging communities, and the limitations of AI looked at.

You all know of random examples, I've seen the AI heat maps for some diagnoses, which don't map pathology at all.

A bit like the what dog is this Husky ID which relied on the presence of snow alone.

There is definitely a place for CAD (computer aided diagnosis) but AIDED is the thing.

There are papers showing reduced reporting time and increased accuracy, but don't forget the "other" bits that make me worthwile - the spots the computer isn't trained on.

Radiologist

[Il'Geller]

Re: and another thing..

The problems that you have identified are not related to AI as such, but to the quality of images that AI uses for its training, as well as the quality of specialists' feedbacks who annotate these images. This Google AI only compares the images and answers questions: it doesn’t think but searches, that’s all it does.

[Il'Geller]

AI can think and offer new ideas, solutions. But in the case of mammograms, the question of liability arises. I don't think the risk will be greater than with a doctor, but...

I have not applied for patents on thinking AI and therefore Google does not know how to make it.