基于深度学习的冠状动脉CT图像分割算法研究  

Research on Coronary CT Image Segmentation Algorithm Based on Deep Learning

作　　者：郭改文 吴笛鸣 王楠 GUO Gaiwen;WU Diming;WANG Nan(School of Computer and Artificial Intelligence,Henan Finance University,Zhengzhou 450046,China;School of Computer Science,China University of Geosciences,Wuhan 430000,China;Henan Silane Technology Development Co.Ltd.,Xuchang 461000,China)

机构地区：[1]河南财政金融学院计算机与人工智能学院,河南郑州450046 [2]中国地质大学计算机学院,湖北武汉430000 [3]河南硅烷科技发展股份有限公司,河南许昌461000

出　　处：《河南财政金融学院学报（自然科学版）》2024年第3期9-13,共5页Journal of Henan Finance University(Natural Science Edition)

基　　金：教育部中国高校产学研创新基金蓝点分布式计算项目(2021LDA11001);河南省科技攻关项目“基于迁移学习的故障诊断方法研究”(232102220022);河南省重点学科计算机科学与技术资助(2023—2027)。

摘　　要：为了更好地协助医生对冠状动脉疾病做出诊断,提升诊断冠状动脉CT图像的分割精度,研究了3种深度学习网络模型算法Transformer、DeepLab V3+和U-Net。首先对冠状动脉图片进行增强、去噪、归一化等预处理操作,以降低分割难度;然后分别利用这3种深度学习网络模型算法对冠状动脉图像进行像素级分割,并对比它们的分割效果。Transformer模型在PA、MPA和MIoU 3个评价指标上均具备最优性能,在医学图像分割精度,尤其是精细边界区域分割方面有较大优势。实验验证,Transformer模型有效可行,同等测试条件下显著优于DeepLab V3+和U-Net模型。To assist doctors in diagnosing coronary artery diseases better and improve the segmentation accuracy of coronary CT images,three deep learning network models,Transformer,DeepLab V3+and U-Net,were studied.First,the coronary images were preprocessed with enhancement,denoising and normalization to reduce segmentation difficulty.Then,three deep learning network models were used to perform pixel-level segmentation of coronary images,and their segmentation effects were compared.The Transformer model achieved the best performance in PA,MPA,and MIoU evaluation metrics,demonstrating significant advantages in the accuracy of medical image segmentation,especially in fine boundary area segmentation.Experiment results show that this method is effective and feasible,significantly outperforming the DeepLab V3+and U-Net models in the same conditions.

关 键 词：冠状动脉疾病 图像分割 TRANSFORMER DeepLab V3+ U-Net 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TP399[自动化与计算机技术—控制科学与工程]