基于Unet+Attention的胸部CT影像支气管分割算法
Research on Trachea Segmentation Algorithm Based on Unet+Attention from Chest CT Images
张子明 1周庆华 2薛洪省 3覃文军1
作者信息
- 1. 东北大学医学影像智能计算教育部重点实验室,沈阳 110189;东北大学计算机科学与工程学院,沈阳 110189
- 2. 东北大学计算机科学与工程学院,沈阳 110189
- 3. 大连大学附属中山医院,辽宁大连 116001
- 折叠
摘要
目前肺气管分割中,由于CT图像灰度分布复杂,分割目标像素近似,易造成过分割;而且肺气管像素较少,难以得到更多目标特征,造成细小肺气管容易被忽略.针对这些难点,本研究提出结合Unet网络和注意力机制的肺气管分割算法,注意力机制使用的是关注通道域和空间域的卷积块注意力模型(CBAM),该模型提高了气管特征权重.在损失函数方面,针对原始数据中正负样本失衡的问题,引入focal loss损失函数,该函数对标准交叉熵损失函数进行了改进,使难分类样本在训练过程中得到更多关注;最后通过八连通域判断将孤立点去除,保留较大的几个连通域,即最后的肺气管部分.选用由合作医院提供的24组CT影像和43组CTA影像,共计26 157张切片图像作为数据集,进行分割实验.结果表明,分割准确率能够达到0.86,过分割率和欠分割率均值为0.28和0.39.经过注意力模块和损失函数的消融实验,在改进前的准确率、过分割率和欠分割率分别为0.81、0.30、0.40,可见其分割效果均不如Unet+Attention方法.与其他常用方法在相同条件下进行比较后,在保证过分割率和欠分割率不变的情况下,所提出的算法得到了最高的准确率,较好地解决了细小气管分割不准确的问题.
Abstract
There are many challenges in current lung trachea segmentation including complex grayscale distribution of CT images,segmentation target pixel approximation,easy to cause over-segmentation,fewer lung trachea pixels,and difficult to get more target features,therefore,fine lung trachea is easy to be ignored.In this paper,we studied a lung trachea segmentation algorithm combining Unet network and attention mechanism,which a convolutional block attention model CBAM focusing on the channel domain and spatial domain was used in the attention mechanism,which improved the tracheal feature weights.In terms of loss function,for the problem of imbalance between positive and negative samples in the original data,this paper used the focal loss function to improve the standard cross-entropy loss function,so that the hard-to-classify samples got more attention in the training process.Finally,the isolated points were removed by eight connected domains judgment,and several larger connected domains were retained,i.e.,the last pulmonary trachea part.Twenty-four sets of CT images and 43 sets of CTA images provided by the partner hospitals,totaling 26 157 slice images,were selected as the data set for segmentation experiments.The results showed that the segmentation accuracy reached 0.86,and the mean values of over-segmentation rate and under-segmentation rate were 0.28 and 0.39 respectively.After the ablation experiments of attention module and loss function,the accuracy,over-segmentation rate and under-segmentation rate before improvement were 0.81,0.30 and 0.40,respectively,indicating the segmentation effect was inferior to the method proposed in this paper.Compared with other commonly used methods under the same conditions,the proposed method reached the highest accuracy rate under the condition that the over-segmentation rate,and under-segmentation rate were guaranteed to be unchanged.The above experiments proved the accuracy of the algorithm in this paper,and successfully solved the problem of inaccurate segmentation of fine trachea.
关键词
医学图像分割/肺气管/Unet/注意力机制/focal/lossKey words
medical image segmentation/lung trachea/Unet/attentional mechanisms/focal loss引用本文复制引用
基金项目
国家自然科学基金(61971118)
中央高校基本科研业务费专项(N2216014)
辽宁省科技计划(2021JH1/10400051)
出版年
2024
NETL
NSTL
万方数据