3D UNeXt:轻量级快速脑提取网络

为了解决现有脑提取网络结构复杂、参数量大且推理速度不高的问题，受UNeXt启发，提出一种基于3D卷积、3D多层感知机（multilayer perception，MLP）和多尺度特征融合的轻量级快速脑提取网络3D UNeXt极大地减少了参数和浮点运算量，取得了令人满意的结果。3D UNeXt以U-Net为基本架构，在编码阶段使用3D卷积模块获取局部特征；在瓶颈阶段通过3D MLP模块获取全局特征和特征之间的远程依赖；在解码阶段借助多尺度特征融合模块高效融合浅层特征和深层特征。特别地，3D MLP模块在三个不同特征轴向进行线性移位操作以获取不同维度特征的全局感受野并建立它们之间的远程依赖。在IBSR、NFBS和HTU-BrainMask三个数据集上进行实验，以和先进网络进行对比。实验结果表明，3D UNeXt在网络参数、浮点运算量、推理精度和速度等方面显著优于现有模型。

3D UNeXt: lightweight and efficient network for effective brain extraction

In order to solve the drawbacks of existing brain extraction network, i. e., complex network structure, large amounts of parameters and low inference speed, this paper proposed a novel network 3D UNeXt for fast and effective brain extraction. 3D UNeXt greatly reduced parameters and the number of floating point operators(FLOPs), and achieved promising results with the combination of 3D convolution, 3D MLP and multi-scale feature fusion. 3D UNeXt used U-Net as the basic architecture and employed 3D convolutional modules to obtain local features in encoding stage. Specifically, the proposed 3D MLP module at the bottleneck stage enhanced the extraction of global features and long-range dependencies among them. In decoding stage, this paper designed a lightweight multiscale feature fusion module to effectively fuse multiscale low-level features and high-level counterparts. In detail, the 3D MLP module performed linear shift operations in three different axes to obtain global receptive fields from different dimension features and establish long-range dependencies among them. This paper computed 3D UNeXt with other counterparts on three datasets, such as IBSR, NFBS, and HTU-BrainMask. Experimental results show that the 3D UNeXt is superior over other baselines in terms of network parameters, FLOPs, inference accuracy, and inference speed.