深度学习实现基于CTA图像的B型主动脉夹层全自动分割
本研究回顾性收集2011年1月至2018年12月276例B型主动脉夹层患者的术前CTA图像,建立三标签手动分割金标准数据库,按约8:1随机分为训练集246例和测试集30例,开发三种卷积神经网络模型。串行多任务模型CNN3表现最佳,全主动脉、真腔和假腔的平均Dice系数分别为0.93±0.01、0.93±0.01和0.91±0.02。深度学习模型为B型主动脉夹层提供了一种有前景的准确高效分割方法,使解剖特征自动测量成为可能。
要点
- 串行多任务CNN3模型在测试集上全主动脉、真腔、假腔平均Dice系数分别为0.93、0.93、0.91。
- CNN3分割速度达0.038±0.006秒/图像,分割体积与金标准一致性良好。
- 深度学习模型使B型主动脉夹层解剖特征自动测量成为可能。
核心结论
- 基于深度学习的模型为B型主动脉夹层提供了有前景的准确高效分割方法,使解剖特征自动测量成为可能。
研究设计与样本
- 回顾性研究,收集2011年1月至2018年12月276例B型主动脉夹层患者的术前CTA图像。
- 采用可重复的三标签手动分割协议(全主动脉、真腔、假腔)建立金标准数据库(n=276)。
- 数据库按约8:1随机分为训练集(n=246)和测试集(n=30)。
- 在训练集上开发三种卷积神经网络模型:单任务CNN1、单多任务CNN2和串行多任务CNN3。
诊断标准与关键数值
- CNN3在测试集上全主动脉、真腔和假腔的平均Dice系数分别为0.93±0.01、0.93±0.01和0.91±0.02(p<0.05)。
- CNN3各标签体积与金标准一致性极好,全主动脉、真腔和假腔的平均体积差分别为-31.05(-82.76至20.65)ml、4.79(-11.04至20.63)ml和8.67(-11.40至28.74)ml。
- CNN3分割速度为0.038±0.006秒/图像。
临床应用与阅片要点
- 原文未据此给出推荐。
局限性
- 原文未报告局限性。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
PURPOSE: This study sought to establish a robust and fully automated Type B aortic dissection (TBAD) segmentation method by leveraging the emerging deep learning techniques.
METHODS: Preoperative CTA images of 276 patients with TBAD were retrospectively collected from January 2011 to December 2018. Using a reproducible manual segmentation protocol of three labels (whole aorta, true lumen (TL), and false lumen (FL)), a ground truth database (n = 276) was established and randomly divided into training and testing sets in a rough 8:1 ratio. Three convolutional neural network (CNN) models were developed on the training set (n = 246): single one-task (CNN1), single multi-task (CNN2), and serial multi-task (CNN3) models. Performance was evaluated using the Dice coefficient score (DCS) and lumen volume accuracy on the testing set (n = 30). Pearson correlation, Intra-class correlation coefficients and Bland-Altman plots were used to evaluate the inter-observer measurement agreement.
RESULTS: CNN3 performed the best, with mean DCSs of 0.93 ± 0.01, 0.93 ± 0.01 and 0.91 ± 0.02 for the whole aorta, TL, and FL, respectively (p < 0.05). Each label volume from CNN3 showed excellent agreement with the ground truth, with mean volume differences of -31.05 (-82.76 to 20.65) ml, 4.79 (-11.04 to 20.63) ml, and 8.67(-11.40 to 28.74) ml for the whole aorta, TL, and FL, respectively. The segmentation speed of CNN3 was 0.038 ± 0.006 s/image.
CONCLUSION: Deep learning-based model provides a promising approach for accurate and efficient segmentation of TBAD and makes it possible for automated measurements of TBAD anatomical features.
引用
- Cao L, Shi R, Ge Y, et al. Fully automatic segmentation of type B aortic dissection from CTA images enabled by deep learning. Eur J Radiol. 2019. PMID: 31683252. doi:10.1016/j.ejrad.2019.108713
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。