AI algorithm capable of diagnosing skin cancer up to 96% correctly

Not only did workers, journalists, and service workers come to the dermatologist but also feared being replaced by machines.

Stanford University researchers have created an AI algorithm that can identify skin cancer as well as a professional doctor. This software is trained with nearly 130,000 images related to moles, blemishes and skin lesions through the form of deep learning.

After that, it "confronted" with 21 dermatologists and obtained the same results as them (at least 91% accurate compared to humans). In the future, they plan to conduct an application to detect skin cancer at home for smartphones.

Each year, there are 5.4 million cases of skin cancer in the United States. The diagnosis process has many stages, including moles or skin marks by specialist doctors. The sooner the disease is detected, the higher the chance of cure and survival.

For example, the early 5-year survival rate of melanoma cell cancer is 97%, but in later stages, the number drops to only 14%.

Melanoma is a very common disease

In order to "teach" their AI how to recognize skin cancer, Stanford scientists used an existing deep learning algorithm , created by Google with image sorting. They described the process in their study: we provided it with thousands of images collected all over the world, along with giving it a picture of what is ill, or is it good evil.

" There is no available database for us to train our algorithms, so we have to create it ourselves," said Brett Kuprel, co-author of the study. " We have to take pictures from the Internet and work with medical schools to create a neat collection from the messy data stack - the name is written in many languages, including German, Arabic and Latin . "

Will this algorithm replace the dermatologist in the future?

The team then created a database of 129,450 images of 2,032 different diseases. The AI ​​system then scans each pixel in these pictures, looking for the similarity of each disease to each of the skin cancer patients. After the training process, the system was able to identify diseases "just as well as trained professionals ," the scientists said.

For example, with melanoma, a dermatologist recognizes 95% of the damage it causes and correctly diagnoses 76% of benign tumors. In that same experiment, the other AI algorithm scored 96% of the skin melanoma and 90% of the harm was not harmful.

Stanford scientists say that the purpose of their software development is not to replace dermatologists, but to help every method of diagnosis not too expensive. They are also aspiring to create an advanced version of this algorithm and turn it into an application that can be used at home.

However, in order for this to happen, the AI ​​needs to practice more (it has become accustomed to working with high-resolution and high-resolution images - not the types of images that smartphones can capture) and need rigorous evaluation in many respects before being released.