摘要
针对海马体图像多图谱分割算法精度低的问题,在多图谱分割的配准环节提出了一种基于U-Net的深度可变形配准模型。将U-Net编码环节的标准卷积替换为深度可分离卷积(DSConv),以增强模型的特征提取能力;引入可变形大核注意力(D-LKA)模块,提高对重要区域特征的注意力;运用空洞卷积(DC)模块扩展感受野,强化对多尺度信息的捕捉能力。改进算法在公开数据集LPBA40与OASIS上的实验结果表明,该模型在OASIS数据集上的配准精度可达0.7988;通过多图谱分割标签融合阶段的多数表决方法,最终分割精度相较于其他配准方法提升了5%~9%。本模型展现了潜在的临床应用价值,在早期阿尔茨海默病诊断中具有积极参考意义。
Aiming at the problem of low accuracy of the hippocampus image multi-atlas segmentation algorithm,a depth deformable registration model based on U-Net is proposed in the registration stage of multi-atlas segmentation.The standard convolution of the U-Net encoding session is replaced with depthwise separable convolution to enhance the feature extraction ability of the model.A deformable large kernel attention(D-LKA)module is introduced to improve the attention to important regional features.By utilizing the dilated convolution module to expand the receptive field,the model strengthens its ability to capture multi-scale information.Experimental results of the proposed algorithm on public available datasets LPBA40 and OASIS show that the model registration accuracy on the OASIS dataset achieves 0.7988,and the final segmentation accuracy is improved by 5%-9%in comparison with other registration methods through the majority voting method in the label fusion stage of multi-atlas segmentation.This model demonstrates potential clinical application value and offers valuable insights in early Alzheimer’s disease diagnosis.
作者
张静
马瑜
巫睿阳
肖博文
ZHANG Jing;MA Yu;WU Ruiyang;XIAO Bowen(School of Electronic and Electrical Engineering,Ningxia University,Yinchuan 750021,China)
基金
国家自然科学基金项目(42361056)
中央支持地方专项资金项目(2023FRD05034)
宁夏重点研发计划高新技术领域项目(2023BDE03002)。
关键词
多图谱分割
海马体
图像配准
标签融合
深度可分离卷积
空洞卷积
multi-atlas segmentation
hippocampus
image registration
label fusion
depthwise separable convolution
dilated convolution
