What would you do if, one day, you suddenly noticed that a mole on your body looked strange? Although this can be a red flag, many people do not go to the hospital in time for a variety of reasons, such as busy work and inconvenient to go to the hospital. Now, artificial intelligence is offering a better solution to this problem: in the future, we might be able to download an APP on our phones, turn on a camera and ask robo-doctors to see if this is an early sign of skin cancer.
Giiso Information, founded in 2013, is a leading technology provider in the field of “artificial intelligence + information” in China, with top technologies in big data mining, intelligent semantics, knowledge mapping and other fields. At the same time, its research and development products include editing robots, writing robots and other artificial intelligence products! With its strong technical strength, the company has received angel round investment at the beginning of its establishment, and received pre-A round investment of $5 million from GSR Venture Capital in August 2015.
In a paper published on the cover of Nature in late January, a team of researchers at Stanford University developed an ARTIFICIAL intelligence that can diagnose skin cancer with the same accuracy as a human doctor. “Dermatologist level Classification of Skin Cancer with Deep Neural Networks.” Using deep learning, they trained the machine to recognize skin cancer symptoms from images of nearly 130,000 moles, rashes and other skin lesions. After comparing the results of 21 dermatologists, they found that the deep neural network’s accuracy was on a level with human doctors, at more than 91 percent.
Deep learning is adding to medicine
Skin cancer is not a prominent member of the cancer family in China because the incidence of skin cancer is lower in the yellow race than in the white race. But skin cancer is one of the most common cancers in the United States. About 5.4 million Americans develop skin cancer each year. Melanoma, for example, has a 97% survival rate if detected and treated at an early stage within five years. But in the later stages, the survival rate drops dramatically to 14 percent. So early screening can mean the difference between life and death for skin cancer patients.
In general, after coming to the hospital or clinic, doctors will conduct clinical screening based on visual diagnosis, followed by dermatoscopy, biopsy and pathological diagnosis of the suspected lesions.
The doctor used a dermatoscope for examination.
But for a variety of reasons, many people don’t make a quick trip to the hospital for minor skin symptoms. As a result, ai-based portable skin cancer diagnosis devices at home will greatly improve screening coverage for early skin cancer and save more lives. But is ai up to the task of screening melanoma from a common mole? The joint research team at Stanford university concluded that deep learning-based robotic doctors are surprisingly accurate in their diagnosis.
“We realized it was possible, that machines could not only do it, they could do it as well as humans,” says Sebastian Thrun, an assistant professor at Stanford’s Artificial Intelligence Lab. “That’s when our thinking changed completely. We said, ‘Look, this is not just a schoolwork, this could be good for humanity’.”
The visual-processing algorithm is based on the current craze for deep learning, which trains machines to perform specific tasks using large amounts of data as examples. It’s not just visual processing that’s been a big hit lately, but other fields as diverse as Alphago, Google’s go AI, which beat World go champion Lee Sedol after learning 30 million pieces of human chess. In machine learning, instead of coding the solution, the computer “fumbles” the solution by learning sample data. In the case of skin cancer diagnosis, the researchers no longer needed to teach the computer the regular features of skin cancer appearance, but the computer itself learned the patterns.
Based on a Google algorithm that distinguishes dogs from cats
Instead of starting from scratch, the developers modelled it on a Google algorithm that can identify 1,000 objects in 1.28 million images. Google’s algorithm was originally designed to distinguish cats from dogs, but now, Researchers trained it to distinguish benign seborrheic keratosis from keratinocyte carcinomas, common moles, and malignant melanomas.
In terms of data, however, the first problem the team faced was that there was no large database of skin cancers available. So the Stanford ARTIFICIAL Intelligence Lab pulled data from the Internet and partnered with the Stanford School of Medicine to categorize and label this jumble of photos. This is not an easy task. After all, there are several languages in the raw data, and unifying these translations would be time-consuming.
The joint team then sifted through the hodgepodge. Professional dermatologists use a dermoscope, a hand-held microscope, to zoom in on the skin in question, producing a medical image with some set criteria. But most of the photos here are not professional medical images, and vary in Angle, size and brightness. In the end, they selected 129,450 images of skin lesions, including 2,032 different diseases. Each image was fed into the algorithm as a pixel with a related disease label. In this way, the developer saves a lot of early image grouping work, greatly improving the amount of data.
Image samples: Benign and malignant epithelial cells/melanocytes/dermoscopic melanocytes.
After training, the researchers tested the machine’s learning using high-quality, biopsy-proven photographs provided by the University of Edinburgh and the International Skin Imaging Collaboration Project, The photos involved two of the most common and deadly types of skin cancer: malignant melanoma and keratinocyte carcinoma. Twenty-one human dermatologists were asked to look at more than 370 of the images and decide for each one whether to perform a further biopsy or treatment, or to tell the patient good news.
In the tests, the AI was asked to perform three diagnostic tasks: identifying keratinocyte carcinoma, identifying melanoma, and classifying melanoma using dermoscopic images. Researchers construct a sensitivity – specificity curve to measure the performance of the algorithm. Sensitivity reflects the algorithm’s ability to correctly identify malignant lesions, and specificity reflects the algorithm’s ability to correctly identify benign lesions, that is, not to be misdiagnosed as cancer. In all three tasks, the AI performed as well as human dermatologists, with a sensitivity of 91 percent.
The sensitivity of the algorithm to diagnose the images of different numbers of keratinocytes and melanocytes is above 91%.
In addition to matching the diagnostic sensitivity of a human doctor, the algorithm’s sensitivity is adjustable. The sensitivity can be adjusted according to the desired diagnostic effect.
Giiso information, founded in 2013, is the first domestic high-tech enterprise focusing on the research and development of intelligent information processing technology and the development and operation of core software for writing robots. At the beginning of its establishment, the company received angel round investment, and in August 2015, GSR Venture Capital received $5 million pre-A round of investment.
The palm doctor of the future
The algorithm still needs to run on a computer, but the Stanford team will try to make it small enough to be loaded on a mobile phone. They found it easy, but needed more clinical testing. In the near future, people may be able to make a reliable diagnosis of skin cancer with the tap of a finger.
Esteva, a graduate student in Thrun’s lab, said, “When I thought about the powerful presence of smartphones, it really clicked for me. Everyone will have a supercomputer in their pocket in the future. What if we use it to screen for skin cancer, for other diseases? “
To be sure, deep learning is a fertile ground for possibilities. The Stanford algorithm for skin cancer screening is just a small opening to a new world where deep learning-based AI will work alongside human doctors in the broader medical field.