TRANSFER LEARNING FOR IMAGE CLASSIFICATION OF PRIMARY MORPHOLOGICAL ELEMENTS OF SKIN LESIONS

• Main Authors: T. A. Polevaya, A. A. Filchenkov, I. A. Saitov, R. A. Ravodin
• Format: Article
• Language: English
• Published: Saint Petersburg National Research University of Information Technologies, Mechanics and Optics (ITMO University) 2019-03-01
• Series: Naučno-tehničeskij Vestnik Informacionnyh Tehnologij, Mehaniki i Optiki
• Subjects: skin disease; primary morphology of skin lesions; transfer learning; machine learning; automatic diagnostics; VGG16
• Online Access: https://ntv.ifmo.ru/file/article/18624.pdf
• ISSN: 2226-1494; 2500-0373

We consider the problem of image classification by deep learning methods for solving classification task for primary morphological elements of skin lesions. The quality of medical care provided to the population depends largely on the medical personnel competence. The problem of medical errors is quite acute in various medical fields, especially, in dermatovenerology. In view of these conditions, the creation of clinical decision support systems becomes one of the promising directions of improving the quality of medical care for patients with dermatovenerological profile. A module of automatic detection of primary morphological elements of skin lesions on skin lesions images can be considered as one of the components of such systems. This study proposes a solution for the problem of primary morphological elements classification based on deep learning and transfer learning. We compare the effect of different learning algorithms application on the accuracy of resulting skin lesion images classifier. We provide experimental results on application of suggested solution to the following primary morphological elements: pustule, macule, nodule, papule and plaque. The proposed algorithm showed 76.00% accuracy for 5 classes of primary morphological elements (pustule, macule, nodule, papule and plaque), 77.50% accuracy for 4 classes (macule, nodule, papule and plaque) and 81.67% accuracy for 3 classes (nodule, papule and plaque).