基于形变残差和级联编码的胰腺分割模型
Pancreas Segmentation Model Based on Deformable Residual and Cascading Encoding
朱纷 1何立风 1孙爽 1张梦颖 1于佳佳1
作者信息
- 1. 陕西科技大学电子信息与人工智能学院,陕西 西安 710021
- 折叠
摘要
为解决深度卷积神经网络进行胰腺分割时存在的胰腺形状位置变化大、噪声干扰、部分小目标等问题,提出一种结合可形变收缩残差块(Deformable Shrinkage Residual Block,DSRB)与级联编码模块(Cascading Encoding Module,CEM)的胰腺分割模型DC U-net.该模型利用2个可形变卷积、注意力机制以及残差结构设计了DSRB,通过可形变卷积来解决胰腺形状位置变化大的问题,使用软阈值化来减少噪声干扰;采用CEM来进行特征融合,对编码特征进行复用以降低编解码阶段的特征差异度,加强对小目标特征的学习.在NIH公开数据集上的实验结果表明,本文模型DC U-net的平均Dice相似系数(Dice Similarity Coefficient,DSC)达到 87.26%,平均交并比(Intersection Over Union,IOU)达到77.98%,分割精度优于对比模型.
Abstract
In order to solve the problems of large pancreatic shape and position change,noise interference,and some small tar-gets in pancreas segmentation by deep convolutional neural networks,a pancreatic segmentation model DC U-net combining de-formable shrinkage residual block(DSRB)and cascading encoding module(CEM)is proposed.The DSRB is designed by using two deformable convolutions,an attention mechanism,and a residual structure.This method solves the problem of large changes in pancreatic shape and position through deformable convolution,and uses soft thresholding to reduce noise interference.CEM is used to fuse features,and the coding features are multiplexed to reduce the feature differentiation in the encoding and decoding stage,and strengthen the learning of small target features.The experimental results on the NIH public dataset show that the pro-posed DC U-net model achieves an average Dice similarity coefficient(DSC)of 87.26%,the average section over union(IOU)of 77.98%,and the segmentation accuracy is better than that of the comparison model.
关键词
图像分割/胰腺分割/可形变收缩残差块/级联编码模块/特征融合Key words
image segmentation/pancreas segmentation/deformable shrinkage residual block/cascading encoding module/feature fusion引用本文复制引用
基金项目
国家自然科学基金资助项目(61971272)
出版年
2024
国家科技期刊平台
NSTL
万方数据