可变形分支注意力融合网络的胰腺分割方法

Pancreas Segmentation Method with Deformable Branch Attention Fusion Network

胰腺具有尺寸小、形状不规则且多变的特点,因此在腹部CT图像中自动分割胰腺具有极大的挑战性.为了适应胰腺特征并解决其分割困难的问题,本文提出了一种轻量级的可变形分支注意力融合网络(Deformable Branch Attention Fusion Network,DBA-Net)作为胰腺自动分割方法.该方法首先将候选区域裁剪出来作为网络的输入,以便减少背景干扰并突出胰腺区域;然后引入可变形卷积使网络自适应地学习胰腺的空间结构;最后提出分支注意力融合模块实现低级别特征和高级别特征的融合,帮助解码器更好地还原特征图.本文的方法在NIH数据集上测试的Dice相似系数为85.3%,在MSD数据集上的Dice相似系数为78.9%,相比基线U-Net分别提高了3.9%和5.6%.实验结果表明本文的方法能够对胰腺进行更好的分割.

The pancreas is characterized by its small size,irregular and variable shape,making the automatic segmentation of the pancreas in abdominal CT images highly challenging.In order to adapt to the unique characteristics of the pancreas and address the difficulties in its segmentation,this paper proposes a lightweight deformable branch attention fusion network(DBA-Net)as an automatic pancreas segmentation method.This method first crops the candidate regions as inputs to the network,reducing background interference and highlighting the pancreatic region.Then,deformable convolutions are introduced to enable the network to adaptively learn the spatial structure of the pancreas.Finally,a branch attention fusion module is proposed to fuse low-level and high-level features,aiding the decoder in better feature map reconstruction.The method proposed in this paper achieved a Dice similarity coefficient of 85.3%on the NIH dataset and 78.9%on the MSD dataset.Compared to the baseline U-Net,it improved the segmentation performance by 3.9%and 5.6%respectively.The experimental results demonstrate that the proposed method achieves better pancreas segmentation.