Group Shift Pointwise Convolution for Volumetric Medical Image Segmentation

Recent studies have witnessed the effectiveness of 3D convolutions on segmenting volumetric medical images. Compared with 2D counterparts, can capture spatial context in three dimensions. Nevertheless, models employing introduce more trainable parameters and are computationally complex, which may lead easily to model overfitting especially for applications limited available training data. This paper aims improve efficiency by introducing a novel Group Shift Pointwise Convolution (GSP-Conv). GSP-Conv simplifies into pointwise ones \(1\times 1\times 1\) kernels, dramatically reduces number FLOPs (e.g. \(27\times \) fewer than \(3\times 3\times 3\) kernels). Naive receptive fields cannot make full use image context. To address this problem, we propose parameter-free operation, (GS), shifts feature maps along different directions an elegant way. With GS, access features from locations, be compensated. We evaluate proposed method two datasets, PROMISE12 BraTS18. Results show that our method, substantially decreased complexity, achieves comparable or even better performance convolutions.