A new AI program trained by an international team of data scientists has demonstrated the ability to best professional dermatologists when it comes to looking at skin lesions like moles and figuring out which ones are benign and which ones are potentially cancerous. The AI was set up as a convolutional neural network and trained by a team consisting of data scientists from the United States, Germany, and France. Once it was all set up, the AI managed to achieve a 95-percent success rate in diagnoses, while a group of select dermatologists was only able to muster 86-percent on average.
The group of dermatologists tapped for the study consisted of a range of experience levels. 29-percent of the dermatologists present had less than two years of experience in the field, about every fifth had two to five years of prior work, and all the rest in the study had at least five years, which meant, for study purposes, that they were considered experts. The AI managed to outperform most of the dermatologists on a one-on-one basis, with only the most seasoned experts able to actually best it. The AI did not err on the side of caution in order to avoid missing diagnoses, either; it was able to detect seemingly benign cancer risks just as well as it was able to mitigate false alarms. In the real world, this would translate to both preventing unnecessary surgeries and being able to catch more patients in an earlier time frame, making treatment easier. The AI will reportedly end up packaged as a consumer-facing tool meant to be a preamble to a doctor visit. The AI may have a hard time catching moles that are in strange places, which means that it still can’t replace a visit to the clinic.
This comes after a Google-built AI fashioned from an existing image recognition algorithm achieved performance in the same type of feat that came close to results from human dermatologists. This exponential growth in skill and accuracy is not building upon the very same AI, but rather representative of advancements in the field of AI in general, and image processing in particular.