In 2017, Stanford scientists created a diagnosis system by analyzing a photo of a patient’s skin. The algorithm is called the “deep convolutional neural network.” It runs on Google Brain, a Google project that aims to explore the possibilities of machine learning. Now, Russian mathematicians have developed a system that recognizes 10 different types of pigmented skin lesions from a photo of a mole.
Identifying the problem
Skin cancer remains one of the most common types of malignant tumors, and is increasingly common due to the intensification of ultraviolet radiation. However, diagnosing this type of cancer is a problem because it is difficult to distinguish a malignant formation from a benign one. Russian scientists from the North Caucasus Federal University decided to turn to artificial intelligence.
They designed a system based on neural networks, which will help increase the accuracy of diagnosis. The team of mathematicians worked with noisy skin images, especially those containing hair. The system they proposed is capable of pre-processing the image to improve the accuracy of detecting melanoma and other pigmented skin lesions.
Replacing hair with pixels
The solution developed involves replacing pixels of hair structures with pixels of the skin. By using this method, it is possible to preserve the diagnostic features in the image. This is a core component of the analysis and the process is divided into a series of stages.
First the RGB-image is decomposed into color components, and then the hair pixels are located and replaced by neighboring pixels. The system then assembles the image back together.
Pigmented skin lesions are then recognized and classified using specially trained AlexNet convolutional neural networks. About 42 thousand clinical dermatoscopic images from the ISIC Melanoma Project are used for this purpose.
As a result, the developed system is able to recognize 10 categories of pigmented skin lesions: from dermatofibroma, nevus, solar lentigo, different types of ketarosis to melanoma and other types of cancer.
The accuracy of the system during testing was 80.81%. This figure is higher than that of other alternatives working on similar methods.
The North Caucasus Federal University team notes that the use of the system will improve the quality of diagnosis and will allow for the beginning of the treatment to start at an earlier stage of the disease. They note that a mobile application will allow for anyone to be able to check themselves for similar skin lesions.
The researchers, similarly, plan to build a more complex system of neural network classification of pigmented skin neoplasms that will use metadata about patients (age, gender, race, genetic predisposition, etc.).
In this manner, it has become clear that neural networks present an opportunity for more accurate skin cancer detection.