基于联合深度网络和形态结构约束的三维医学图像分割方法  

3D Medical Image Segmentation Based on Joined Depth Network and Morphological Structure Constraints

作　　者：李军 叶欣怡 杨长才[1] 陈秋凤 薛岚燕[1] 魏丽芳[1] Li Jun;Ye Xinyi;Yang Changcai;Chen Qiufeng;Xue Lanyan;Wei Lifang(College of Computer and Information,Fujian Agriculture and Forestry University,Fuzhou 350002)

机构地区：[1]福建农林大学计算机与信息学院,福州350002

出　　处：《中国生物医学工程学报》2023年第1期30-40,共11页Chinese Journal of Biomedical Engineering

基　　金：国家自然科学基金(62171130);福建省自然科学基金(2020J01573)。

摘　　要：医学图像自动分割具有广泛和重要临床应用价值,特别是病灶、脏器的自动分割。基于传统图像处理方法的医学图像分割仅能利用浅层结构模型的浅层特征来识别感兴趣区域,并且需要大量人工干预。而基于机器学习的分割方法在模型建模时存在局限性且缺乏可解释性。本研究提出一种基于Transformer和卷积神经网络结合形态结构约束的三维医学图像分割方法。编码器中利用卷积神经网络和Transformer构建U型网络结构提取多重特征;解码器中采用上采样并通过跳跃连接将不同层次的特征拼接在一起;加入形态结构约束模块,通过提取病灶和脏器等分割目标的形状信息,以增强模型可解释性,并采用最大池化和平均池化操作,对经过卷积神经网络得到的结果进一步提取有代表性的特征,作为形态结构模块的输入,最终提高分割结果的准确性。在公开数据集Synapse和ACDC上利用评价指标Dice相似系数(DSC)和Hausdorff距离(HD)验证所提出算法的有效性。其中,在Synapse数据集上,18例数据作为训练集,12例数据作为测试集;在ACDC数据集上,70例数据作为训练集,10例数据作为验证集,20例数据作为测试集。实验结果表明,在Synapse上使用不同的优化器的Dice相似系数(DSC)和HD均值可以达到76.67%和25.18 mm(SGD优化器)及82.80%和21.07 mm(Adam优化器);在ACDC上不同的优化器的DSC可以达到90.65%(SGD优化器)及91.75%(Adam优化器),与其他方法相比具有一定的优势。所提出的方法对三维医学图像分割中的错分割问题有所改善,并提高了图像分割任务的性能。Automatic segmentation of medical images has extensive and important clinical application value,especially the automatic segmentation of lesions and organs.The medical image segmentation based on conventional image processing methods can only utilize shallow features extracted by shallow structure model to identify the regions of interest and requires a lot of manual intervention.However,the segmentation methods based on the machine learning have limitations and lack of interpretability in modeling.This paper presented a 3D medical image segmentation method based on Transformer and convolutional neural network(CNN)combined with morphological structure constraints.In the encoder,the CNN and Transformer were used to construct a U-shaped network structure to extract various features;and in the decoder,the up-sampling operation was used and the features of different levels were concatenated together by skip-connections.The morphological structure constraint module was added to enhance the interpretability of the model through extracting the shape information of segmented targets such as lesions and organs,and the maximum pooling and average pooling operations were used to further extract representative features from the results obtained through the CNN as the input of the morphological structure module and improved the accuracy of the final segmentation results.The evaluation indexes DSC and HD were used to verify the effectiveness of the proposed algorithm on the public datasets Synapse and ACDC.On the Synapse dataset,18 cases of data were used as the training set and 12 cases of data were used as the test set;on the ACDC dataset,70 cases of data were used as the training set,10 cases of data were used as the validation set and 20 cases of data were used as the test set.The experimental results showed that the average value of DSC and HD of different optimizers reached 76.67%and 25.18 mm(SDG)and 82.80%and 21.07 mm(Adam)on Synapse respectively,and the average value of DSC and HD of different optimizers reached 90.65%

关 键 词：三维医学图像 图像分割 TRANSFORMER 卷积神经网络 形态结构约束 

分 类 号：R318[医药卫生—生物医学工程]