Lin Wei-Wei, Liu Tian-Jian, Dai Wen-Li, Wang Qiang-Wei, Hu Xin-Ben, Gu Zhao-Wen, Zhu Yong-Jian
Department of Neurosurgery, The Second Affiliated Hospital of Zhejiang University School of Medicine, Hangzhou, China.
School of Mathematical Sciences, Zhejiang University, Yuquan Campus, Lingyin Street, Hangzhou, 310013 Zhejiang, China.
Eur J Radiol. 2022 Jun;151:110287. doi: 10.1016/j.ejrad.2022.110287. Epub 2022 Apr 2.
This study aimed to evaluate the diagnostic performance of convolutional neural network (CNN) models in Chiari malformation type I (CMI) and to verify whether CNNs can identify the morphological features of the craniocervical junction region between patients with CMI and healthy controls (HCs). To date, numerous indicators based on manual measurements are used for the diagnosis of CMI. However, the corresponding postoperative efficacy and prognostic evaluations have remained inconsistent. From a diagnostic perspective, CNN models may be used to explore the relationship between the clinical features and image morphological parameters.
This study included a total of 148 patients diagnosed with CMI at our institution and 205 HCs were included. T1-weighted sagittal magnetic resonance imaging (MRI) images were used for the analysis. A total of 220 and 355 slices were acquired from 98 patients with CMI and 155 HCs, respectively, to train and validate the CNN models. In addition, median sagittal images obtained from 50 patients with CMI and 50 HCs were selected to test the models. We applied original cervical MRI images (CI) and images of posterior cranial fossa and craniocervical junction area (CVI) to train the CI- and CVI-based CNN models. Transfer learning and data augmentation were used for model construction and each model was retrained 10 times.
Both the CI- and CVI-based CNN models achieved high diagnostic accuracy. In the validation dataset, the models had diagnostic accuracy of 100% and 97% (p = 0.005), sensitivity of 100% and 98% (p = 0.016), and specificity of 100% (p = 0.929), respectively. In the test dataset, the accuracy was 97% and 96% (p = 0.25), sensitivity was 97% and 92% (p = 0.109), and specificity was 100% (p = 0.123), respectively. For patients with cerebellar subungual herniation less than 5 mm, three out of the 10 CVI-based retrained models reached 100% sensitivity.
Our results revealed that the CNN models demonstrated excellent diagnostic performance for CMI. The models had higher sensitivity than the application of cerebellar tonsillar herniation alone and could identify features in the posterior cranial fossa and craniocervical junction area of patients. Our preliminary experiments provided a feasible method for the diagnosis and study of CMI using CNN models. However, further studies are needed to identify the morphologic characteristics of patients with different clinical outcomes, as well as patients who may benefit from surgery.
本研究旨在评估卷积神经网络（CNN）模型在I型Chiari畸形（CMI）中的诊断性能，并验证CNN是否能够识别CMI患者与健康对照（HC）之间颅颈交界区的形态特征。迄今为止，基于手动测量的众多指标被用于CMI的诊断。然而，相应的术后疗效和预后评估仍不一致。从诊断角度来看，CNN模型可用于探索临床特征与图像形态参数之间的关系。
本研究共纳入了在我们机构诊断为CMI的148例患者，并纳入了205例HC。使用T1加权矢状位磁共振成像（MRI）图像进行分析。分别从98例CMI患者和155例HC中获取了总共220层和355层图像，用于训练和验证CNN模型。此外，选择了从50例CMI患者和50例HC中获得的正中矢状位图像来测试模型。我们应用原始颈椎MRI图像（CI）以及后颅窝和颅颈交界区图像（CVI）来训练基于CI和CVI的CNN模型。使用迁移学习和数据增强进行模型构建，每个模型重新训练10次。
基于CI和CVI的CNN模型均取得了较高的诊断准确率。在验证数据集中，模型的诊断准确率分别为100%和97%（p = 0.005），灵敏度分别为100%和98%（p = 0.016），特异性均为100%（p = 0.929）。在测试数据集中，准确率分别为97%和96%（p = 0.25），灵敏度分别为97%和92%（p = 0.109），特异性均为100%（p = 0.123）。对于小脑扁桃体下疝小于5毫米的患者，基于CVI重新训练的10个模型中有3个达到了100%的灵敏度。
我们的结果表明，CNN模型在CMI诊断中表现出优异的性能。这些模型比仅应用小脑扁桃体下疝具有更高的灵敏度，并且能够识别患者后颅窝和颅颈交界区的特征。我们的初步实验为使用CNN模型诊断和研究CMI提供了一种可行的方法。然而，需要进一步研究以确定具有不同临床结局的患者以及可能从手术中获益的患者的形态学特征。
