Crowdsourced Medical Imaging Analysis

Every so often, people need to undergo one of the many medical imaging procedures. There are numerous techniques, from the common X-ray for bone fractures, to magnetic resonance imaging for abnormalities of soft tissue.

In the past, X-rays for example, relied upon large sheets of photographic film which were looked at by a busy junior doctor at the end of a 30hr shift, They looked at the suspected area before declaring the metacarpal broken and the person a fool for hitting a wall. Nowadays, X-rays can be electronically displayed. Whereupon the suspected area is looked at, the metacarpal declared broken, and the person declared a fool for hitting a wall. More advanced imaging is handled in a similar way, although there is often an imaging technician available to help interpretation.

What links diverse medical imaging techniques, is that they're often looking for something specific, like a broken metacarpal or evidence of congenital abnormalities. Once the tired junior doctor has confirmed a broken bone, they're likely to carry on along their original plan in the absence of any fragments or other related complications. The electronic image is then filed away on a hard drive somewhere until it is no longer needed, accidentally deleted or distributed to the world as part of a colossal data breach during the latest big budget IT failure.

As someone who spends a tremendous amount of time analyzing imaging data, I'm acutely aware of how much I'm guided by what I am looking for. Often, quite glaringly obvious phenomena are missed because you're simply not looking for it. Alternatively, you may have learned that spending half a day looking at B is somewhat detrimental to your detailed analysis of A, so its consciously or unconsciously ignored. A third way, is that you display the images in such a way as to maximize your chances of noticing phenomenon A, while B disappears into the noise.

I suspect similar things happen with medical imaging all the time. Its easy to miss a grape seed sized tumour in the pancreas when you're a hepatologist looking for gall stones.

If Amazon's Mechanical Turk is anything to go by, it's surprisingly easy to get large numbers of people to perform computer based tasks for little money. This is the model that could be used.

Once electronic imaging data is collected, the data should be carefully stripped* of identifying data and uploaded to a server. From here the imaging data may be analyzed by freelance people from around the world. Any oddities within the imaging data are then referred up to people with some actual training. Initially, much dreadful dross will be flagged up. Eventually however, individuals will emerge with real talent, either generally or in specific types of imaging. These people can gradually be promoted, with progressive financial incentives. Soon they'll be spotting real medical issues within the images and these can be referred all the way along to the managing physician. Often, these will simply confirm the problem that the imaging was performed for. But every now and then, they'll spot something the medical staff missed, especially if several people flag the same thing in the data.

The best people to set this up would be the companies that make the equipment. The financial incentives would likely be easily covered by the hospital who will be happy to take the extra business**. The companies can boast about enhanced detection rates. Strangely, people will also do it out of curiosity, like with the planet-finding, or out of some macabre fascination with broken metacarpals. Perhaps people will try and cheat, using automated analysis, but still, good luck to them.

*carefully, not like normal. ** assuming the US system of hospitals and insurance companies looking for extra justification to screw each other