Reinventing 2D Convolutions for 3D Images

There have been considerable debates over 2D and 3D representation learning on medical images. approaches could benefit from large-scale pretraining, whereas they are generally weak in capturing large contexts. natively strong contexts, however few publicly available dataset is diverse enough for universal pretraining. Even hybrid (2D + 3D) approaches, the intrinsic disadvantages within / parts still exist. In this study, we bridge gap between convolutions by reinventing convolutions. We propose ACS (axial-coronal-sagittal) to perform learning, while utilizing pretrained weights datasets. convolutions, convolution kernels split channel into three parts, convoluted separately views (axial, coronal sagittal) of representations. Theoretically, ANY CNN (ResNet, DenseNet, or DeepLab) able be converted a CNN, with weight same parameter size. Extensive experiments several benchmarks (including classification, segmentation detection tasks) validate consistent superiority CNNs, counterparts without can used as plug-and-play replacement standard convolution, smaller model size less computation.