Earlier this month, Venture Beat revealed that Google developed an artificial intelligence (AI) program that has the potential to be a revolutionary MedTech development. The firm’s system can detect abnormal 26 skin conditions. Moreover, the corporation published a study stating that its deep learning system has an accuracy rate comparable to a U.S. board-certified dermatologist.
A Technological Solution
According to the Global Research on the Impact of Dermatological Diseases project, 25 percent of medical treatments provided around the world are for skin conditions. Due to the scope of this problem, most sufferers seek treatment from a general practitioner as opposed to a specialist. As a result, well-meaning clinicians often misdiagnose patients’ dermatological ailments.
Researchers at Google sought to address this problem by developing a deep learning solution capable of identifying different dermatological conditions. Using a combination of images and patient metadata, engineers developed an artificial intelligence (AI) program that can accurately detect 26 skin conditions. Notably, the application behind the system was able to replicate a critical medical tool: differential diagnosis.
When identifying a skin problem, dermatologists make a list of potential ailments that could be responsible for a patient’s condition. Then, they use lab tests, imaging, and other procedures to whittle down the list to make a definitive diagnosis. Google’s AI enacts the same process by cross-examining a database of dermatological ailments with a wealth of patient metadata.
Indeed, researchers found that they could train their deep learning solution to a high degree of accuracy.
Impressive Accuracy Rates
Google’s data scientists used a vast quantity of data to train their AI. The programmers fed their system information from 17,777 identified cases from 17 primary care clinics spread across two states. Moreover, the researchers also provided their system with 50,000 differential diagnoses performed by 40 dermatologists. When evaluating their system, researchers found that the AI identified dermatological ailments with increasing degrees of accuracy based on the number of images and patient metadata with which it was supplied. Indeed, when provided with six images, the program could detect skin conditions as well as a board-certified dermatologist.
Furthermore, with an accuracy rate of 71 to 91 percent, Google’s diagnostic AI proved more effective at identifying dermatological illness than general practitioners (60 percent) or nurse practitioners (55 percent).
Notably, the corporation’s deep learning system did not have the racial misidentification issues that plague other AI recognition programs. Google’s AI had an accuracy rate of 69 to 94 percent among individuals with pale white to dark brown skin.
Despite their success, the programmers behind the diagnostic AI note that it does have some significant limitations. As an example, the researchers couldn’t program their AI to identify certain rare skin tones because they lacked sufficient data. Similarly, the system had difficulty identifying conditions like melanoma due to a lack of sample data. However, the deep learning team is optimistic that their system could help improve dermatological diagnosis with further development.
For instance, the Google data scientists note that their system could eventually help clinicians without dermatological expertise perform more accurate triage. Moreover, the group believes its AI could serve as a second opinion and help specialists expedite their diagnostic process. Either way, the diagnostic deep learning solution could improve the standard of care for billions of people.