Medical Image Segmentation Used Unsupervised Convolutional Neural Network

Abstract

Segmenting skin lesions is a crucial step in computer-aided melanoma diagnosis, but it is also a highly difficult task because of the imprecise lesion boundaries and varied lesion textures. We describe a deep fully convolutional network-based, totally automatic technique for skin lesion segmentation (FCNs). In order to extend ConvNets for tasks requiring medical picture segmentation, this work provides unsupervised domain adaption techniques employing adversarial learning. By combining the prior knowledge stored by the shallow network with the deep FCNs, we are able to show that our newly built network may increase accuracy for lesion segmentation. Our approach is robust and does not require extensive parameter tuning or data augmentation. The experimental findings demonstrate the method's considerable potential for efficient model generalization.