AI and the science of medical imaging

Sir, – Charlotte Blease asserts in her article that AI can “read cancer scans quicker, more consistently and more accurately than the average human doctor” (“AI: better at reading scans and a better bedside manner too?”, Science, Opinion, October15th). This is incorrect, at least for now.

At present AI is only useful for answering specific questions in medical imaging. One AI programme widely used in Irish hospitals can flag the presence of a particular type of brain haemorrhage on a CT scan. When this same programme is shown a different type of haemorrhage, it does not detect it. In a similar vein, different AIs are available to detect lung cancer on chest X-rays and breast cancer on mammograms. But these algorithms do not allow for other diagnoses such as a punctured lung or mastitis.

Indeed as Hannah Fry notes in her 2019 book Hello World, a team of trained pigeons have been shown to beat human radiologists at identifying breast cancer on mammograms. But I have yet to encounter a pigeon working in a hospital.

To say AI (or a pigeon) “reads” scans is inaccurate. To “read” a scan involves more than just confirming or refuting one diagnosis. It requires us to interpret the whole scan in the context of the patient’s symptoms and other test results, and to consider the full range of possible diagnoses. As yet this is beyond the capability of AI. – Is mise,