AI is becoming a major tool for people in many fields, especially in healthcare. A new deep learning network called deep learning convolutional neural network (CNN) has been developed to help dermatologists. Through testing, the new CNN has been proven to detect skin cancer even better than expert dermatologists. The CNN was developed by researchers from around in the world, including Germany, the United States and France. The new development is important because 232,000 new malignant melanoma cases are found around the world every year and 55,500 people die from melanoma cases each year. If the CNN is developed to be used as a tool in a dermatologist’s office, many lives could be saved.
The researchers showed the CNN 100,000 images of both malignant melanomas and benign moles. The system read the images and learned how to differentiate harmless moles from cancerous spots. After learning, testing showed that the CNN missed fewer melanomas and misdiagnosed fewer moles than the dermatologists in the study.
To test the CNN, 58 dermatologists from 17 countries volunteered to participate in the study. The real dermatologists had a range of experience with dermoscopy. Seventeen of the dermatologists had less than two years of experience, 11 dermatologists had two to five years of experience and 30 dermatologists had 5 or more years of experience.
There were two levels of testing that the dermatologists participated in during the study. In the first level of the study, the dermatologists detected around 86 percent of the melanomas and identified 71.3 percent of the benign lesions. In the second level of testing, the dermatologists were able to diagnose 88.9 percent of the malignant tumors and 75.7 percent of the non-cancerous legions. In the first level of testing, the expert dermatologists performed better than the less experienced dermatologists, but both groups still didn’t perform better than the CNN.
The research team doesn’t want to develop the CNN to replace dermatologists completely. They want the CNN to work hand in hand with the doctors to diagnose cancerous spots and hopefully save people from cancer with early detection.
While the study is promising, there are some limitations to the testing. The dermatologists who participated in the study were in an artificial setting where they knew they were being tested. Because of this, the results may be different for the test than in real life. The test also didn’t include a full range of skin lesions, so the CNN has only learned the details about melanomas. Most of the skin images that were shown to the CNN were white and had limited genetic backgrounds, so the CNN has a lot of learning to do in order to be used for everyone in a doctor’s office setting. The researchers also say that doctors might not want to use a CNN in their office because they don’t trust the artificial intelligence.
There are a few things that need to be improved with the system before it is implemented in doctor’s offices. Currently, the system can only study melanomas in obvious places. The system has not been able to accurately diagnose moles that are in strange places, like on the head. It also doesn’t know how to recognize atypical melanomas.
"Currently, there is no substitute for a thorough clinical examination," the authors said. "However, 2D and 3D total body photography is able to capture about 90 to 95 percent of the skin surface and given the exponential development of imaging technology we envisage that sooner than later, the automated diagnosis will change the diagnostic paradigm in dermatology. Still, there is much more work to be done to implement this exciting technology safely into routine clinical care."
The paper on this research and the new system was published in Annals of Oncology.
