| Abstract [eng] |
In this research, the accuracy of deep learning models is investigated. The object of analysis is a derivative of skin cancer - melanoma. Image recognition, segmentation is used. Input data are medical photos. The photos are taken from the one of the largest publicly available database – ISIC18. In the research work theoretical analysis of deep learning models is performed. Related works on this topic and their results are reviewed. After reviewing the theoretical part, the research and improvement of the accuracy of the U-net model begins. The original architecture of the U-net model is chosen, which is supplemented in the workflow with blocks that improve accuracy, also the structure of the network decoding part is changed. 3 different model configurations are used to know which structure of the model leads to the highest accuracy. Duration of the deep learning process 50 iterations. 1815 images used for training, the test sample consists of 520 images, validation - 259. The highest achieved result according to the F1 evaluation criterion is 89,21 percent. Using one of the best-performing structures, multiclass segmentation is tested, during which classification is performed - 5 classes of derivatives are distinguished from 1 image. |
