基于多模态MRI与卷积神经网络的较低级别胶质瘤IDH突变预测研究

Accurately assessment of isocitrate dehydrogenase (IDH) mutation is crucial to clinical decision making and prognosis assessment of lower-grade gliomas (LrGG, Grade II/III) patients. However, it may be difficult for existing detection methods to determine IDH status of LrGG patients before clinical intervention. Studies indicate that magnetic resonance imaging (MRI) with different modalities have the potential of predicting IDH mutation of LrGG patients preoperatively and noninvasively. Furthermore, the development of deep learning in medical imaging field recently, make image-based IDH mutation prediction more feasible. On the basis of previous experiment of our group, this research plan to explore a multi-channels deep convolutional neural network (CNN) for multi-modal MRI of LrGG patients, and trained the multi-channels CNN model using transfer learning and layer-wise fine-tuning strategy, which can solve the CNN training problem caused by small sample image dataset; compare the discrimination capacity of IDH mutation between CNN-generative feature and handcraft feature that used in previous study, and finally verify the superiority of CNN model on IDH prediction performance. This research is promising to propose an noninvasive technique for preoperative IDH mutation prediction and provide quantitative indicator for clinical decision, and therefore, has important research significance and application value.

异柠檬酸脱氢酶（IDH）基因突变的准确评估，对较低级别胶质瘤（LrGG，II/III级胶质瘤）患者的手术、辅助治疗方案制定与预后评估意义重大，然而现有检测手段无法在临床干预前确定LrGG患者的IDH状态。有研究表明，不同模态磁共振影像（MRI）在不同层面具有术前无创预测LrGG患者IDH突变的潜能。且近年来，深度学习在影像诊断领域的兴起进一步为基于影像的IDH突变预测创造了可能。本项目拟基于前期研究，探索适用于LrGG患者多模态MRI的多通道深度卷积神经网络（CNN）结构，并利用迁移学习与逐层微调策略解决基于小样本的CNN训练难题，最终建立具有IDH突变预测能力的CNN模型；比较CNN模型的生成特征与传统人工特征的IDH突变鉴别能力，验证CNN模型的IDH突变预测效能优越性。本研究有望为LrGG患者提供一种术前无创IDH突变预测技术，为临床决策提供定量依据，具有重要研究意义和应用价值。

异柠檬酸脱氢酶（IDH）基因突变的准确评估，对较低级别胶质瘤（LrGG，II/III级胶质瘤）患者的手术、辅助治疗方案制定与预后评估意义重大，然而现有检测手段无法在临床干预前确定LrGG患者的IDH状态。有研究表明，不同模态磁共振影像（MRI）在不同层面具有术前无创预测LrGG患者IDH突变的潜能。且近年来，深度学习在影像诊断领域的兴起进一步为基于影像的IDH突变预测创造了可能。本项目基于前期研究，探索适用于LrGG患者多模态MRI的多通道深度卷积神经网络（CNN）结构，并利用迁移学习与逐层微调策略解决基于小样本的CNN训练难题，最终建立具有IDH突变预测能力的CNN模型；比较CNN模型的生成特征与传统人工特征的IDH突变鉴别能力，验证CNN模型的IDH突变预测效能优越性。本研究有望为LrGG患者提供一种术前无创IDH突变预测技术，为临床决策提供定量依据，具有重要研究意义和应用价值。