It may come as a shock, but computers are not color blind, literally and figuratively. Much like in society, AI shows a racial preference towards light-skinned people. But, before you start throwing around blame, it is not the computer's fault. This preference stems from a lack of diverse training data.
Computer Vision & Dermatology
In recent years, interest in using computer vision, a subset of AI, to diagnose skin diseases has grown exponentially. Computer vision aims to train computers to interpret and understand the visual world like the human eye. They accomplish this by using digital images from cameras and videos and deep learning models to identify and classify objects. Dermatologists hope to use computer vision to mainstream the diagnosis process through mobile apps. Their goal is to create an algorithm that can diagnose skin diseases using a photo of the skin taken by the patient. But there is one problem that continues to challenge developers: creating an algorithm that can accurately diagnose both light-skinned and dark-skinned patients.
AI Melanoma Detection
Melanoma, a type of skin cancer, is commonly characterized by the darkening of the skin. For the human pupil, it isn’t hard to notice a change in skin color, but for an algorithm, it can be a challenge. It has to determine the natural skin color from the image, after considering light exposure and camera focus. For people with paler skin, it is easier to differentiate the skin from the blemish because there is more contrast in color. That means, even with less diverse training data, the algorithm will still be able to yield accurate results because minor changes in light or focus won't affect the diagnosis. But for darker skin tones, a tiny increase in brightness can make all the difference.
Why Does it Matter?
Currently, the amount of data on dark-skinned people used to train the algorithms is limited. Yet, they are the group that requires the most data for the algorithms to work accurately for them. It is up to us to make sure the data includes the necessary training data to produce equally accurate results for both light and dark-skinned people. We need to recognize that if we want AI to imitate human behavior, then the computer is going to see skin color just as we do. In the current climate of our nation, we must take it upon ourselves to create AI that can treat patients equally well regardless of race. By developing medical technology that is accessible to and accurately treats all patients, we are doing our part to combat racial injustice.
Commenti