Siddhartha Mukherjee points to problems in Deep Learning

During his keynote address at the recent RSNA conference in Chicago, Dr. Siddhartha Mukherjee – an oncologist and Pulitzer Prize-winning author of The Emperor of all Maladies – identified artificial intelligence as a key development for medicine. However, he also emphasized that doctors don’t quite trust AI, and that there’s a fair degree of skepticism about it.

“There’s a hesitancy to embrace AI technology because of the limitations of the technology – AI networks still remain black boxes,” said Dr. Mukherjee, addressing the Radiological Society of North America. “Physicians are trained mechanistically, in why something works some way. Most of the algorithms offer no perspective on why a particular lesion is classified as benign or malignant. They only offer the end-point solution. There’s hesitancy because our brains are trained to think about why.”

He illustrated this “doctor’s dilemma” by pointing to a dermatology project that employed AI to detect cancerous skin lesions.

The researchers put together challenge sets of hundreds of thousands of lesions and used them to train algorithms. “To the human eye, it is difficult or even impossible to tell the difference between some of the lesions,” he noted.

Indeed, in the images he displayed on a large screen, some of the lesions that looked benign were actually malignant, while others that looked malignant were benign.

However, the AI-powered system that was created performed well on the large test set, and efficiently separated malignant skin tumours from the benign lesions.

Upon further analysis, the researchers discovered that the in the training set, some radiologists had marked the lesions they thought were malignant with a yellow mark. “So that’s what the algorithm was picking up,” said Dr. Mukherjee. “It’s a case of garbage in, garbage out.”

That’s a kind of worst-case scenario that can disturb and dismay physicians about AI.

Dr. Mukherjee spent some time on the underpinnings of AI. When it comes to Deep Learning, he asked, “How do we learn? Can machines learn like us? And can machines learn medicine?”

He pointed to the philosopher Gilbert Ryle, who made the distinction between “knowing that” and “knowing how”.

“Knowing about bicycles, for example, doesn’t mean you know how to ride one,” said Dr. Mukherjee. “Think about how you learnt to ride a bicycle, or how you taught your children to ride. What you didn’t do is hand them a manual that said, climb on the seat (step one), push with 60 psi on the front pedal (step two), push your back foot with 10 psi (step three), balance yourself and hold the handlebars at the same time.”

What you did instead was show your kids how to ride the bike.

Encapsulating the “knowing-how” into computer AI systems may be something of a roadblock now, as the solutions tend to the “knowing that” side in their algorithms.

Gilbert Ryle was very interested in the difference between rule-based or algorithmic learning and experience-based learning. And he said, “Rules, like birds, must live before they can be stuffed.”

He added, “That’s going to become very important, because as we move forward, we’ll understand what these lived rules are.”