ARE WE JUST ‘PRACTICING’ MEDICINE?
Healthcare today has become more of “practice of medicine” rather than “science of medicine.”
Take for example, the standard treatment for fever has remained the same for decades.
The same goes for diagnosis of cancers. Pathologists have been largely diagnosing in more or less the same way for the past 100 years. They do so by laboring manually over a microscope reviewing biopsy samples. This is almost identical to working robotically, sifting through thousands of cells before identifying few diseased ones.
The task is often tedious, error-prone and increases subjectivity.
When something as a diagnosis of cancer is the hallmark of the “practice of medicine” and hasn’t been challenged for decades, we have to ask: Can artificial intelligence compete with doctors?
MACHINE LEARNING: SOURCE OF HEALTHCARE INNOVATION?
Scientists and engineers, working at Stanford University certainly feel so. The research team, from the electrical engineering and dermatology department have trained a computer to detect and differentiate between different types of skin cancer.
Guess what it performed as well as the board-certified dermatologists it was tested against. Not one but all 21 of them.
The original idea behind this was from software developed by Google. The AI had learned to spot the difference between images of cats and dogs.
Andre Esteva, PhD student and first author of the study says, “[The computer] could identify 200 breeds of dogs. So why not use it for something more useful.”
The team then programmed the computer for pattern recognition on millions of images, which took about a week. At first, they subjected the AI to healthy skin and then to images of the specific non-malignant lesion.
The algorithm, potentially based on traits like contrast and coloration, learned to differentiate between healthy skin and lesion. They added in and fine-tuned the machine learning on about 130,000 clinical images of 2,032 skin diseases, which the computer mastered overnight.
It then learned to detect the hallmarks and draw conclusions about the most deadly and most common type of skin cancer: melanoma and carcinoma. Now thats some cool machine learning.
BUT HOW DOES THE ALGORITHM RANK AGAINST HUMAN DERMATOLOGISTS?
Quite well. In one trial, the AI scored 71% accuracy against the humans’ average score of 66%. In another trial, published in Nature, it outperformed a panel of 21 dermatologists. Not only this, the team tasked the AI with analyzing the lesion image and diagnose, whether to treat or reassure the patient.
Guess what, again the AI was more accurate than most doctors, but under-performing in just two scenarios.
The picture looks rosy, isn’t it? There is a caveat, though. The study does not mention the most obvious character: skin color. All the images feature people with lighter skin color.
As a result, it gets easier to differentiate between normal skin and malignant lesions. Just as dreadful this is, this will shock you further. Previous algorithms tend to favor whites and shockingly have labeled black people as gorillas.
We certainly don’t want that. Researchers should make it more robust by training the algorithm to work on people with all shades of skin.
Sorry folks, the algorithm is only useful for a segment of the global population, at least for now.
CAN WE HAVE A SMARTPHONE APP, PLEASE!
Yes!. The research team is interested in pursuing a smartphone app. But for now, the algorithm only runs on computers. The team feels it will be relatively easy to develop a smartphone app. Remember the app, skin vision.
The algorithm for this is based on irregularities in color, texture, and shape of the lesion. Maybe we can have more sound proof diagnosis based on machine learning.
Brett Kuprel says, “It is not meant to replace dermatologists. It’s a tool that can assist them.” Susan Swetter, director of Pigmented Lesion and Melanoma Program at Stanford Cancer Institute feels “Advances in computer-aided diagnosis could provide better management options for skin cancer patients. But, before implementation in clinical practice, this needs rigorous validation globally.”
Even though the algorithm is at infancy and machine learning has a long way to go, researchers feel that someday technology will take over 80% of what doctors do.
Imagine your smartphone diagnosing skin cancer. Scary enough!!!