Akademik Veri Yönetim Sistemi
Biomedical Signal Processing and Control, cilt.85, 2023 (SCI-Expanded)
Early and accurate diagnosis plays a significant role in increasing the survival rate of patients with skin cancer. Therefore, automated methods with convolutional neural networks have been developed to help dermatologists in making their decision about the disease. However, these kinds of networks have these two important shortcomings that cause low performance when real-word images are used: (i) They need huge number of training datasets, and (ii) they lead to the loss of information because of the pooling process. Although, capsule networks have been proposed recently to solve these issues and to achieve classification of skin cancers with high accuracy, they have their own defects or disadvantages. Therefore, in this work, a new neural network constructed by adjustable and convolutional capsule layers is proposed. In the adjustable capsule layer, the spatial relationships of capsule vectors are encoded with a constant learnable bias. The capsule vectors are used in a sub-network to obtain adjustable values. The proposed network can learn encoding and modelling of the spatial relation between capsule vectors and preserves the vectors’ orientations. The main contributions of this work are as follows: (i) A novel network including capsule layers is presented. (ii) The performance of the proposed network in the multi-class classification of skin cancer is presented. (iii) Comparative evaluations of the capsule networks applied for the classification of skin cancers are presented. Experiments indicate that the proposed network is promising in the classification of seven kinds of skin cancer (with an accuracy of 95.24%).