Infant brain MRI segmentation with dilated convolution pyramid downsampling and self-attention

TL;DR

This paper introduces a dual aggregation network with dilated convolution pyramid downsampling and self-attention modules to improve infant brain MRI segmentation, achieving state-of-the-art results in DICE scores.

Contribution

It proposes a novel dual aggregation network with specific modules to better preserve spatial details and enhance feature representation in infant brain MRI segmentation.

Findings

DICE ratio of WM and GM increased by 0.7% over previous methods.

Achieved first place in DICE for WM and GM in the iseg-2019 challenge.

CSF segmentation performance is comparable to top methods.

Abstract

In this paper, we propose a dual aggregation network to adaptively aggregate different information in infant brain MRI segmentation. More precisely, we added two modules based on 3D-UNet to better model information at different levels and locations. The dilated convolution pyramid downsampling module is mainly to solve the problem of loss of spatial information on the downsampling process, and it can effectively save details while reducing the resolution. The self-attention module can integrate the remote dependence on the feature maps in two dimensions of spatial and channel, effectively improving the representation ability and discriminating ability of the model. Our results are compared to the winners of iseg2017's first evaluation, the DICE ratio of WM and GM increased by 0.7%, and CSF is comparable.In the latest evaluation of the iseg-2019 cross-dataset challenge,we achieve the first place in the DICE of WM and GM, and the DICE of CSF is second.