LambdaUNet: 2.5D Stroke Lesion Segmentation of Diffusion-Weighted MR Images

Diffusion-weighted (DW) magnetic resonance imaging is essential for the diagnosis and treatment of ischemic stroke. DW images (DWIs) are usually acquired in multi-slice settings where lesion areas two consecutive 2D slices highly discontinuous due to large slice thickness sometimes even gaps. Therefore, although DWIs contain rich 3D information, they cannot be treated as regular or images. Instead, somewhere in-between (or 2.5D) volumetric nature but inter-slice discontinuities. Thus, it not ideal apply most existing segmentation methods designed either To tackle this problem, we propose a new neural network architecture tailored segmenting 2.5D data such DWIs. Our network, termed LambdaUNet, extends UNet by replacing convolutional layers with our proposed Lambda+ layers. In particular, transform both intra-slice context around pixel into linear functions, called lambdas, which then applied produce informative features. LambdaUNet simple yet effective combining sparse information from adjacent while also capturing dense contextual features within single slice. Experiments on unique clinical dataset demonstrate that outperforms 3D/2D image including recent variants UNet. Code available. (URL: https://github.com/YanglanOu/LambdaUNet.)