Mathematicians of the NCFU have developed a system that can recognize pigmented skin neoplasms. The system relies on a photo of a skin area with a mole to identify 10 different types of pigmented skin lesions, while the accuracy of the outcome is above those obtained from similar methods or through visual examination by a specialist.

Researchers note that skin cancer is one of the most common malignancies, and the rate is growing year after year, which is due to the UV-radiation. Diagnosing the dangerous health issue is not always easy, so artificial intelligence may serve good support to medical experts here. Automated recognition systems are already being developed all over the world, while the system proposed by the NCFU experts features certain advantages.

– We have studied similar approaches and could see they fail to be highly efficient, – Ulyana Lyakhova (Head of Project, Postgraduate student, research fellow of Department for Mathematical Modeling) noted. – This largely happens due to interferences on the image, especially such as hair. This creates an occlusion and may have a serious effect on the size, the color and the shape of the skin lesion. To resolve the issue, we decided to have a preliminary image processing thus enhancing the reliability and the detection rate.

Once processed, the image is subject to recognition and restoration. However, what NCFU’s experts have proposed, implies substituting the image areas showing hairs with pixels featuring skin, this maintaining the diagnostic accuracy.

The entire process involves training for neural networks, whereas such training was based on 42K clinical images of skin from the ISIC Melanoma Project (Most of the images are digitized slides of the Roffendaal Clinic for the Treatment of Skin Cancer in Queensland, as well as the Department of Dermatology of the Medical University of Vienna). Now, the system developed in the NCFU is able to detect 10 categories of skin lesion.

Another idea is to develop a mobile app so that anyone can check their own skin and, if necessary, take respective measures contacting a medical specialist.

Future plans include developing more complex systems for neural classifications in view of factors like the patient’s age, sex, race, and genetic predisposition.

For more details, please read the article in the latest issue of Computer Optics (Vol. 45(5) DOI: 10.18287/2412-6179-CO-863, p. 728-735).