Recently, Nature magazine published an article on the cover: using deep learning algorithm to diagnose skin cancer, the accuracy reached 91%, can be compared with doctors.
I wonder if you remember how Google Neural Network distinguishes cats from dogs? Artificial intelligence is different from human beings. A child knows what a cat looks like after meeting it several times. But for a machine, people need to feed it tens of thousands of pictures before it can recognize what a cat is.
Similarly, if people provide high-quality pictures of skin cancer to AI systems, the system can also identify skin cancer by machine learning. Recently, Stanford University published a related research in Nature, and compared the AI system with 24 senior dermatologists. The accuracy of the system is about 91%.
Andre Esteva, the first author of the article and a graduate student at Stanford University, said: "We have developed a very powerful artificial intelligence algorithm that can learn from data and let the system find what to identify and find by itself by writing code."
This algorithm is called convolutional neural network. It first appeared in the brain of Google. It can enhance the decision-making ability of the algorithm by using its amazing computing power. After Stanford University's research, neural networks have been able to recognize 1.28 million images from about 1,000 different categories, but researchers need to understand malignant tumors from a benign seborrheic keratosis.
It is not harmful to distinguish a dog from a pile of Persian cats, but how to distinguish the different spots of various skin diseases and identify skin cancer from them is a matter of life and requires high accuracy.
Colour spots on skin are a big problem. How to distinguish them is a difficult problem to screen image data.
Filtering image data
Brett Kuprel, co-author of the article, a Stanford graduate student, said: "Another difficulty of the study was that there was not enough high-quality skin cancer image data to train artificial intelligence algorithms, and we had to solve it ourselves."
Even before dealing with pictures, they have to do some translation work: "We searched some pictures on the Internet, cooperated with medical schools, sorted and labeled these pictures - these labels include German, Arabic, Latin and so on."
It is necessary not only to translate and organize, but also to process images. Dermatologists often use an instrument called dermoscope to examine patients carefully. Therefore, medical personnel diagnose diseases by medical images with similar magnification and perspective.
But the pictures of the Internet vary greatly. Some are mobile phones, some are instruments, some are cameras, and the effects of different environments are different. The angles, focal lengths and lighting aspects are also different.
Finally, the researchers collected about 13000 images of skin lesions, including more than 2,000 different diseases. They use these images to create an image library and provide it to the algorithm as the original pixel, each with a label describing the additional data of related diseases. Then the researchers developed an algorithm to figure out the intrinsic link between these pictures: the rules governing the appearance of disease spread through tissues.