Method of cleaning hair structures for intellectual image classification of skin neoplasms

Abstract

The main problem in the application of artificial intelligence in the field of dermatology is the low level of accuracy of recognition and classification systems. This paper proposes a method for identifying and cleaning hair, as well as evaluating the effectiveness of the method using a pre-trained neural network. When examining pigmented lesions, the presence of hair in images can obscure important diagnostic information, reducing the effectiveness and quality of examination results. This method allows the preparation of dermatoscopic images for further automated classification and diagnosis of pigmented skin lesions. Modeling was carried out using the MatLab R2019b software package on clinical dermatoscopic images from the international open archive ISIC Melanoma Project. The proposed method made it possible to achieve the accuracy of recognition of images of pigmented skin lesions in 10 diagnostically important categories up to 80.81%. The use of neural network classification systems for dermatoscopic images of pigmented skin formations with a preliminary stage of hair removal will allow specialists to increase the efficiency and speed of diagnosis and start treatment of the disease at an earlier stage