From the Departments of Radiology and Nuclear Medicine (H.v.V., P.R.K., L.M.v.P., B.J.E., C.B.L.M.M., H.A.M.)
Biomedical Engineering and Physics (H.v.V., P.R.K., L.M.v.P., M.W.A.C., H.A.M.).
AJNR Am J Neuroradiol. 2022 Aug;43(8):1107-1114. doi: 10.3174/ajnr.A7582. Epub 2022 Jul 28.
Supervised deep learning is the state-of-the-art method for stroke lesion segmentation on NCCT. Supervised methods require manual lesion annotations for model development, while unsupervised deep learning methods such as generative adversarial networks do not. The aim of this study was to develop and evaluate a generative adversarial network to segment infarct and hemorrhagic stroke lesions on follow-up NCCT scans.
Training data consisted of 820 patients with baseline and follow-up NCCT from 3 Dutch acute ischemic stroke trials. A generative adversarial network was optimized to transform a follow-up scan with a lesion to a generated baseline scan without a lesion by generating a difference map that was subtracted from the follow-up scan. The generated difference map was used to automatically extract lesion segmentations. Segmentation of primary hemorrhagic lesions, hemorrhagic transformation of ischemic stroke, and 24-hour and 1-week follow-up infarct lesions were evaluated relative to expert annotations with the Dice similarity coefficient, Bland-Altman analysis, and intraclass correlation coefficient.
The median Dice similarity coefficient was 0.31 (interquartile range, 0.08-0.59) and 0.59 (interquartile range, 0.29-0.74) for the 24-hour and 1-week infarct lesions, respectively. A much lower Dice similarity coefficient was measured for hemorrhagic transformation (median, 0.02; interquartile range, 0-0.14) and primary hemorrhage lesions (median, 0.08; interquartile range, 0.01-0.35). Predicted lesion volume and the intraclass correlation coefficient were good for the 24-hour (bias, 3 mL; limits of agreement, -64-59 mL; intraclass correlation coefficient, 0.83; 95% CI, 0.78-0.88) and excellent for the 1-week (bias, -4 m; limits of agreement,-66-58 mL; intraclass correlation coefficient, 0.90; 95% CI, 0.83-0.93) follow-up infarct lesions.
An unsupervised generative adversarial network can be used to obtain automated infarct lesion segmentations with a moderate Dice similarity coefficient and good volumetric correspondence.
监督深度学习是 CT 扫描脑梗死病灶分割的最新方法。监督方法需要手动对病灶进行标注，以便为模型开发提供数据，而生成式对抗网络等无监督深度学习方法则不需要手动标注。本研究旨在开发并评估一种生成式对抗网络，以对随访 CT 扫描的梗死和出血性卒中病灶进行分割。
训练数据来自 3 项荷兰急性缺血性卒中介入治疗试验的 820 例患者的基线和随访 CT 扫描。优化生成式对抗网络，通过生成差值图（从随访 CT 扫描中减去该差值图），使带有病灶的随访 CT 扫描转换为不带病灶的生成基线 CT 扫描。利用生成的差值图自动提取病灶分割。将原发性出血性病变、缺血性卒中出血性转化及 24 小时和 1 周随访梗死病变的分割结果与专家标注进行比较，采用 Dice 相似系数、Bland-Altman 分析和组内相关系数进行评估。
24 小时和 1 周随访梗死病变的中位数 Dice 相似系数分别为 0.31（四分位距，0.08-0.59）和 0.59（四分位距，0.29-0.74）。出血性转化（中位数，0.02；四分位距，0-0.14）和原发性出血病变（中位数，0.08；四分位距，0.01-0.35）的 Dice 相似系数低得多。24 小时（偏倚，3 mL；一致性界限，-64-59 mL；组内相关系数，0.83；95%CI，0.78-0.88）和 1 周（偏倚，-4 mL；一致性界限，-66-58 mL；组内相关系数，0.90；95%CI，0.83-0.93）随访梗死病变的预测病变体积和组内相关系数均良好。
无监督生成式对抗网络可用于获得中等 Dice 相似系数和良好体积一致性的自动梗死病灶分割。
