Curious to know what sort of methods you used then if you don't mind sharing.

The best results were with backpropagation neural networks: http://www.sciencedirect.com/science/article/pii/0888613X949...

But we also used fuzzy logic neural networks with genetic algorithms: http://ieeexplore.ieee.org/document/712156/?reload=true

It were 24 years ago.

There might be some interesting things that can be learned from this kind of info and applied to the current status quo (I'm definitely not arguing that there is a sociopolitical element).

Maybe if MRI scans get cheap enough (due to advances in cheap superconductors or whatever) that it's economically feasible to scan people regularly as a precautionary measure (rather than in response to some symptom), then the bulk of the cost might then be in having the radiologist look at the scans. In those "there's nothing wrong but lets check anyways" cases, it might be better to just have the AI do it all even if its accuracy is lower, if it represents a better health-care-dollar-spent to probability-of-detecting-a-serious-problem ratio. (If the alternative is to just not do the scan because the radiologist's fees are too expensive, then it's better to have the cheap scan than nothing at all.)

I can see an argument that if the company was sued then it could try to push the blame onto the software vendor, but surely that would be decided based on the contract between company and software vendor, which is usually defined by the software license.

Machine learning is already used in Radiology. Chances are eventually Radiology will be the domain of machines. But it's going to take some time to get there. Healthcare is extremely regulated and closed minded.

Most of the people in the thread you listed above are clearly biased towards medicine and against computer science and machine learning. But machine learning has been having success in diagnostic medicine even well before the deep learning boom that thread talks about.

For example: http://ieeexplore.ieee.org/abstract/document/1617200/

- Some RETS recorded requests/responses: https://github.com/estately/rets/blob/master/test/vcr_casset...

(Basically some XML-based (SOAP?), cookie + authorization, that seems very ASP.net / Windows Server centric)

- DICOM ("It includes a file format definition and a network communications protocol"): https://en.wikipedia.org/wiki/DICOM

It's basically how to imaging devices communicate and store the images.

Image examples: http://www.osirix-viewer.com/resources/dicom-image-library/

A video of what a doctor would see: https://www.youtube.com/watch?v=Prb5lcR8Jqw

TCP-based protocol in Wireshark: https://wiki.wireshark.org/Protocols/dicom

I wrote a little .dcm to .jpg based on ruby-dicom BTW: https://gist.github.com/Dorian/9e3eb5891b49926c15a05c641ffef...

- PACS seems just like a database model basically http://mehmetsen80.github.io/EasyPACS/

It's the server that is gonna give the info to the doctors.

It's interesting how there seems to be only one popular viewer: OsiriX

Main issue is with HL7 is not technical. From a business point of view, the incentive to cooperate with other systems via HL7 means another reason for a department to adopt a system other than yours.

These are examples of next-generation radiology companies. The current generation of products are focused on image storage and display. These new companies offer automated image analysis before the radiologist even looks at the image. iSchemaView does hemorrhage maps as soon as new head CT or head MRI is acquired.

It looks like everybody sitting on their data is hindering progress. Is there anything that can be done about that politically? I mean, in many cases the data belongs to the public anyway, unless people signed a waiver, but what is the legality of that?

I'm sure machines will someday take over radiology but there will be many, many jobs automated before it (i.e. decades).

There are three areas that take a lot of time that radiologist would like to see automated:

1. Counting lung nodules. 2. Working mammography CAD. 3. Automated bone-age determination.

Those are the hot three topics for machine learning. Personally, I think that a normal vs. non-normal classifier for CXRs would be more interesting because you could have a completely generated note for normal reads, and radiologists could just quickly look at the image without writing/dictating anything. Of note, hospitals and radiology departments typically lose money on X-ray reads because the reimbursement is $7-$20 (compared to $100+ for MR/CT). So if you could halve the read time, they might become profitable again.

Edit: In terms of 10x, what you'd want is a system that would automatically make the reads (i.e. radiologist report), and a very efficient way for radiologist to verify what is written. It's hard to make a pathologic read, but since roughly 50% of reads are normal, you could start with normal reports.

And then bringing checklist driven analysis for radiologist.

So I decided to observe an radiologist at an hospital for a day (back in 2011), and I noticed that most if not all of it was already automated.

Radiologists were there to rubber-stamp the machine's work (and to ensure compliance, laws and what not).