Exploiting Cross-Dimensional Dependency Using Vision-LSTM for Efficient Medical Image Segmentation

Modeling inter-dependency among various channels or spatial locations is a prevalent method to segment relevant regions in the input. However, capturing cross-dimensional dependency facilitates understanding intricate relationships among various axes. This leads to understanding the complex structure of target regions that may often spread over multiple slices of volumetric medical images. This necessity drives the development of the novel Rotational U- Vision-LSTM (Rot-UViL), for medical image segmentation. An input tensor is rotated across multiple branches to capture the correlation among different dimensions using Vision-LSTM blocks. The outputs from multiple branches are combined to integrate the inter-dimensional and spatial information. The proposed framework exhibits superior performance in terms of different metrics. It provides higher resource efficiency, as compared to related popular baseline algorithms. The code is available at: https://github.com/duttapallabi2907/Rot-UVil.git.