Skin Lesion Classification by Ensembles of Deep Convolutional Networks and Regularly Spaced Shifting

Skin lesions are caused due to multiple factors, like allergies, infections, exposition the sun, etc. These skin diseases have become a challenge in medical diagnosis visual similarities, where image classification is an essential task achieve adequate diagnostic of different lesions. Melanoma one best-known types vast majority cancer deaths. In this work, we propose ensemble improved convolutional neural networks combined with test-time regularly spaced shifting technique for lesion classification. The builds several versions test input image, which shifted by displacement vectors that lie on regular lattice plane possible shifts. subsequently passed each classifiers ensemble. Finally, all outputs from yield final result. Experiment results show significant improvement well-known HAM10000 dataset terms accuracy and F-score. particular, it demonstrated our combination ensembles yields better performance than any two methods when applied alone.