Skin Cancer Diagnosis based on Improved Multiattention Convolutional Neural Network

Abstract

The human skin is directly exposed to air and sunlight, because the skin is the body’s first line of defense against harmful substances, so it may lead to skin cancer due to a variety of reasons. The malignant tumors on the skin are called skin cancers, which are defined by different types of tumor cells, including epidermis, soft skin tissue, melanocytes, skin lymphatic reticulation and hematopoietic tissue, etc. As for the current diagnosis, skin cancer can be diagnosed as a tumor of the skin. Therefore, it is extremely important to develop a diagnostic system for skin cancer. Convolutional neural network has excellent performance in medical image recognition, so this paper proposes an improved convolutional neural network for skin cancer diagnosis, which has the following advantages: (1) by introducing a hybrid multi_attentive mechanism will, the limited attention will be focused on the tumor cells, thus saving resources and obtaining effective eigenvalues quickly. (2) by using the improved vgg19 to extract the eigenvalues of tumors, the fully connected layer is replaced by extremely randomized trees for classification. The study shows that the diagnostic method in this paper has some improvements in roc, accuracy, and recall.