A 3D Shrinking-and-Expanding Module with Channel Attention for Efficient Deep Learning-Based Super-Resolution

Abstract

The 3-dimensional (3D) super-resolution (SR) for medical volumetric data is confirmed to provide better visual results compared to conventional 2-dimensional processing. Then, considering practical applications, we propose a novel convolutional neural network for a 3D SR. Moreover, to reduce model parameters and achieve high precision, we reformulating the standard convolution layer by shrinking the input feature dimension before mapping and expanding back afterward using pointwise convolution. Thus, we utilize channel attention to optimize the network for optimum performance. Although the proposed architecture can be embedded into any 3D model, it performs better while combined with the state-of-the-art model, the 3D residual dense network (RDN). Furthermore, we demonstrate that our proposed method outperforms the state-of-the-art methods in terms of accuracy with only approximately 1/8 parameters and 3/5 times of the standard model, 3D RDN. Consequently, our network is quite suitable in practical situations, such as small training samples, short processing time, and being embedded into chips.