The layers of the Earth make each region unique, just as each human body is unlike any other, according to the vice president of AI at Aidoc, Idan Bassuk. This is the reason that when the special IDF unit digs test-tunnels like those dug in the Gaza Strip, it needs to test its technology under similar conditions to those Hamas worked under.
Top of his 2006 Talpiot class, Bassuk is more than an AI expert. He served in a combat role and led men tasked with facing the tunnels threat.
“This gave me intimate knowledge of not just the reality on the ground,” he says, “but what a soldier actually needs in his hands.”
Unlike sonar waves that “cross” a homogeneous substance, sea-water, to unveil what is hidden in the bowls of the earth requires something like an X-ray. Just as the X-ray presents an image of the inner organs concealed under muscle and fat, the IDF is able to produce an image of what is buried deeper than a metro station.
Just as soil north of the Gaza Strip isn’t like the soil in its south, different hospitals use different machines and protocols to provide different medical images. This might seem odd, but there is no single protocol for medical scanning across the country or the larger medical community. Different hospitals use different machines, which require different manuals and training. “In some cases it comes down to the skills of the technician using the machine,” Bassuk explains.
This poses a question, how do you know the scan you are holding, of the Gaza Strip or the patient, is a good one? You only have one patient, or one tunnel, in the real world. So you only have one piece of data. What can you compare it with?
Take spinal injuries, for example. These are fairly rare and during one month a single hospital might have several dozen examples of scans showing them. But thousands are needed to improve the AI capabilities. Using the solution Aidoc created, it is possible to take thousands of such scans from other countries and “present” the data as if it originated in the hospital based on the population it usually serves. This in turns, allows to AI to “learn” from a small set of cases as if it had a wide pool of many spinal injury scans. Similar means were used to “multiply” the tunnel-situations the AI was “scanning” to enhance its abilities to best respond to the singular one in the factual world.
To use an analogy, in the real world one patient gets one scan, a single roll of the dice, but thanks to the virtual world, a patient can get an AI which learned from thousands of possible scans, thousands of dices rolled elsewhere, after they were adapted for such a purpose. These virtual results aren’t used in the medical care giving process, only to check how well the AI is able to function on a wider set of cases.
In the army, checking thousands of virtual tunnels allows to see where the system needs to work harder to offer unique solutions in specific points, in medical care, better AI means more efficient treatments to the singular patient seeking help. Improved, and early, detection of cancer tumors for example, could save countless lives.
“All these technologies, MRI, X-Rays, are very challenging to us because it’s hard to tell causation and correlation apart,” Bassuk concludes, “the technology offers data, only rigorous methodology can ensure solid understanding.”
How does one “clean” the data received from the machines so one clearly sees where tunnels, or life threatening tumors, are gathered?
One part of the answer is an AI solution engineer. A person who can help the AI correctly “filter” the data regardless of the various variants involved.
“Our goal is to do the same thing for hospitals,” Bassuk says, “not to replace the radiologist – they are the best at understanding human variants – but to put a human in the data processing loop.”