SUNet++: A Deep Network with Channel Attention for Small-Scale Object Segmentation on 3D Medical Images

As a deep learning network with an encoder-decoder architecture, UNet and its series of improved versions have been widely used in medical image segmentation great applications. However, when to segment targets 3D images such as magnetic resonance imaging (MRI), computed tomography (CT), these models do not model the relevance vertical space, resulting poor accurate analysis consecutive slices same patient. On other hand, large amount detail lost during encoding process makes incapable segmenting small-scale tumor targets. Aiming at scene target images, fully new neural SUNet++ is proposed on basis UNet++. improves existing mainly three aspects: 1) modeling strategy slice superposition thoroughly excavate dimensional information data; 2) by adding attention mechanism decoding process, small scale picture are retained amplified; 3) up-sampling transposed convolution operation further enhance effect model. In order verify model, we collected produced dataset hyperintensity MRI liver-stage containing over 400 cases liver nodules. Experimental results both public proprietary datasets demonstrate superiority three-dimensional images.