Accurate MR image super-resolution via lightweight lateral inhibition network

Abstract

In recent years, convolutional neural networks (CNNs) have shown their advantages on MR image super-resolution (SR) tasks. Many current SR models, however, have heavy demands on computation and memory, which are not friendly to magnetic resonance imaging (MRI) where computing resource is usually constrained. On the other hand, a basic consideration in most MRI experiments is how to reduce scanning time to improve patient comfort and reduce motion artifacts. In this work, we ease the problem by presenting an effective and lightweight model that supports fast training and accurate SR inference. The proposed network is inspired by the lateral inhibition mechanism, which assumes that there exist inhibitory effects between adjacent neurons. The backbone of our network consists of several lateral inhibition blocks, where the inhibitory effect is explicitly implemented by a battery of cascaded local inhibition units. When model scale is small, explicitly inhibiting feature activations is expected to further explore model representational capacity. For more effective feature extraction, several parallel dilated convolutions are also used to extract shallow features directly from the input image. Extensive experiments on typical MR images demonstrate that our lateral inhibition network (LIN) achieves better SR performance than other lightweight models with similar model scale.