Long-Sequential-U-Net

Abstract

Given the semantic segmentation problem of 3D medical body scan images, it is necessary to grasp not only the correlation between pixels in a 2D frame but also the sequential contexts between frames so that the pixel-labeling can be more accurately performed. In this thesis, we adopt Convolutional LSTMs to consider the sequential contexts and propose a U-Net-based architecture which includes them. We have studied how to maximize the effects of convolutional LSTMs while minimizing the number of trainable parameters. Compared to the basic U-Net and the previous studies that take into account the sequential contexts, our model achieves high dice scores when solving multi-organ segmentation tasks of SegTHOR dataset.