首页|融合CNN和Transformer的颅内动脉瘤CTA图像分割

融合CNN和Transformer的颅内动脉瘤CTA图像分割

扫码查看
原文链接
  • 万方数据
  • 维普
目的 探讨融合CNN和Transformer的深度学习模型在颅内动脉瘤CTA图像分割中的应用效果。方法 回顾性收集108例未破裂颅内动脉瘤患者的CTA影像数据，分为训练集88例、测试集20例。采用融合CNN-Transformer混合结构的nnFormer模型对CTA图像中的颅内动脉瘤进行自动分割。针对颅内动脉瘤存在位置分布不均、边缘模糊等特点，对模型损失函数进行优化，使用Dice相似系数值、平均交并比(mIoU)、精确率和召回率作为评价指标来评价模型的分割效果，以医生人工标记为参考标准。结果 在测试集上的Dice相似系数值、mIoU、精确率和召回率分别达到了0。842、0。739、0。844、0。861，在颅内动脉瘤分割任务中取得了较为优异的结果。与其他分割方法相比，各项评价指标均有不同程度的提升。结论 融合CNN和Transformer的深度学习模型能对颅内动脉瘤CTA图像进行精准分割，有效提高医生的诊断效率，临床应用价值较高。
Intracranial Aneurysm CTA Image Segmentation by Fusing CNN and Transformer
Objective The application effect of the deep learning model fusing CNN and Transformer in CTA image segmentation of intra-cranial aneurysm was discussed. Methods CTA imaging data of 108 patients with unruptured intracranial aneurysms were retrospec-tively collected,and they were divided into 88 cases in the training set and 20 cases in the test set. The nnFormer model fused with CNN-Transformer hybrid structure was used to automatically segment intracranial aneurysms in cranial CTA. In view of the uneven location dis-tribution and fuzzy edges of intracranial aneurysms,the loss function of the model was optimized,and the Dice similarity coefficient value,average intersection union ratio(mIoU),precision and recall rate were used as evaluation indicators to evaluate the segmentation effect of the model,and the doctor's manual marking was used as the reference standard. Results The Dice similarity coefficient,mIoU,accu-racy,and recall of this method on the test set reached 0.842,0.739,0.844,and 0.861,achieving excellent results in intracranial aneurysm segmentation. Compared with other segmentation methods,all evaluation indicators have been improved to varying degrees. Conclusion The deep learning model by fusing CNN and Transformer can accurately segment CTA images of intracranial aneurysms,effectively im-proving the diagnostic efficiency of doctors and having good clinical application value.

intracranial aneurysmloss functionmedical image segmentationdeep learning

陈璇、张雪原、王家琦、殷鹏展、叶明全

展开 >

皖南医学院 医学信息学院,安徽 芜湖 241002

皖南医学院第一附属医院 放射科,安徽 芜湖 241001

颅内动脉瘤 损失函数 医学图像分割 深度学习

安徽省重点研究与开发计划项目安徽省高校协同创新项目安徽省高校优秀科研创新团队项目安徽省高校学科(专业)拔尖人才学术资助项目基金安徽省高校自然科学研究重大项目

2022a05020011GXXT-2022-0442022AH010075gxbjZD20220422023AH040251

2024

吉林医药学院学报
吉林医药学院

吉林医药学院学报

影响因子：0.459
ISSN：1673-2995
年,卷(期)：2024.45(5)