AI model built on Intel processors helps diagnose throat cancer

Artificial intelligence cancer detection is making its way out of hospitals and into patients’ pockets with a new model developed by researchers in Taiwan and the US.

The proof-of-concept diagnostic model aims to pick on cases of laryngeal cancer – a type of throat cancer – simply by listening to a person’s voice for a few seconds. Laryngeal cancer tends to affect the quality of a person’s voice, often making it sound hoarse.

The app listens to the sound of a patient making the noise ‘aah’ for a few seconds. The AI then does its work, comparing the sound against pathological and non-pathological characteristics to detect which voices are more likely to be suggestive of laryngeal cancer. The model is 80% accurate, Intel said.

The link between changes to the sound of a person’s voice and cancer isn’t new. Researchers have shown that simply listening for changes to your speech can not only tell you there’s a problem, but also indicate what type of problem – from potentially life-threatening cancer to benign polyps on the vocal cords.

According to Intel, researchers have been working on the question of how to use voice data to make a reliable model for laryngeal cancer detection for over two decades. However, researchers needed help training an AI to crunch all the vast reams of voice data available and develop a model that can predict laryngeal cancer with some certainty.

A team of hospitals and research institutions, including Far Eastern Memorial Hospital, Taichung Veterans General Hospital in Taiwan and Vanderbilt University Medical Center in the US, were involved in the project.

The app uses Intel software, including Intel’s OpenVINO as well as OpenFL, and runs on 4th Generation Xeon Sapphire Rapids processors and FPGA accelerators.

Faster and cheaper screenings of throat cancer

Having an app that can potentially detect laryngeal cancer is interesting for a number of reasons. If doctors want to rule in or rule out laryngeal cancer, they’ll typically first need to arrange for some form of endoscopy (where a camera is put down a patient’s nose or throat) or arrange imaging such as a CT scan.

Needless to say, tests like these are expensive, time-consuming and not much fun for the patient. Where GPs want to refer their patients for such tests or for further specialist assessment, they currently do so based on symptoms alone. An AI app like this one could potentially be used in a GP’s office as a first-line screening test to find those patients in need of the most urgent referrals, sparing some low-risk patients from invasive tests and fast-tracking other, higher-risk individuals.

Like most cancers, the earlier that laryngeal cancer is detected, the better the prognosis. However, in the UK at least, laryngeal cancer isn’t something that’s routinely screened for. Having an app that can pick up people with an increased likelihood of cancer in a matter of seconds opens up the possibility of opportunistic screening – setting up a pop-up booth and asking people at events such as sports matches to say ‘aah’ into an app could potentially find cases that would otherwise have gone undiscovered until symptoms progressed.

And lastly, there’s the possibility of opening up testing in rural and low-resource environments. For those whose GP or specialist centre is too far or too expensive to access, the ability to quickly screen for laryngeal cancer could be incredibly useful in picking up cancer in its early stages and signposting patients for further investigations.