GEMO:结合几何与形态特征的深度学习脑纤维分类方法
本研究提出一种名为GEMO的监督深度学习脑纤维分类方法。该方法将纤维流线从三维坐标空间通过xyz2RGB映射转换为二维彩色编码图像,输入卷积神经网络,并融合流线的几何与形态特征。结果显示,加入这些特征可提高分类准确率,最高达98.34%,且在六项分割准确性指标上优于现有方法。
要点
- GEMO仅依赖纤维流线自身特征即可完成全脑纤维分类,无需结构像或预处理。
- 将三维流线转为二维彩色编码图像后,可由常规卷积神经网络处理。
- 结合几何与形态特征较单纯深度学习特征能提高纤维分类准确率。
核心结论
- GEMO方法仅利用纤维流线自身特征即可完成全脑纤维束分类,无需结构图像或预处理。
- 纳入流线几何与形态特征可提高深度学习网络的纤维分类准确率,最高达98.34%。
- GEMO在六项分割准确性指标上均优于现有方法,可提供自动化白质束分割。
研究设计与样本
- 研究提出一种监督深度学习纤维分类方法,称为GEMO。
- 采用xyz2RGB映射方法将三维流线坐标转换为二维彩色编码图像,输入卷积神经网络进行学习与分类。
- 除卷积神经网络特征外,同时提取流线的几何特征与形态特征并纳入分类。
关键影像表现/征象
- 流线经xyz2RGB映射后的二维彩色编码图像可作为卷积神经网络输入。
- 单独分析几何与形态特征时,可将网络准确率提高至98.34%。
诊断标准与关键数值
- GEMO在全脑纤维束分割的六项准确性指标上表现优于其他现有方法。
- 几何与形态特征单独分析时可将网络准确率提高至98.34%。
临床应用与阅片要点
- 该方法可对全脑纤维束图像进行自动化白质束分割,为结构连接评估和神经外科计划提供帮助。
- GEMO不需要结构图像或预处理,便于在缺少结构像的条件下使用。
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
RATIONALE AND OBJECTIVES: Diffusion-weighted magnetic resonance imaging (dMRI) is often used to study brain structure. One of the important applications made possible by dMRI is streamline tractography that is utilized for structural connectivity evaluation and neurosurgical planning. Reliable localization of brain fibers can considerably enhance the outcome of these applications. In this study, a supervised deep learning approach entitled "GEMO" was proposed for fiber classification.
MATERIALS AND METHODS: This approach benefits from GEometrical and MOrphological features in addition to the features extracted from a convolutional neural network for improving the final classification performance. Streamlines are transformed from three-dimensional coordinate space into two-dimensional color-encoded images using the "xyz2RGB" mapping method and are subsequently fed into a convolutional neural network for learning and classification.
RESULTS: Our results demonstrate that incorporating morphological and geometric features of streamlines enhances the accuracy of the deep learning network in the classification of fibers in comparison with the case where no extracted features were employed. An individual analysis of these features further revealed that they can increase the network's accuracy to as high as 98.34%. GEMO demonstrated consistently higher performance across six segmentation accuracy metrics compared with other existing approaches.
CONCLUSION: Our method performs the streamline classification from a whole brain tractogram, only focusing on each streamline's features, which include those provided from CNN in addition to geometric and morphologic features. GEMO does not need any structural images or preprocessing and outperforms the state-of-the-art segmentation methods. Moreover, it can facilitate further assessment of brain structures by providing automated white matter tract segmentations.
引用
- Barati Shoorche, Makkiabadi, Farnia, et al. GEMO: A Deep Learning Method for Brain Fiber Classification and Tract Segmentation Using Geometrical and Morphological Features. Acad Radiol. 2026. PMID: 42538220. doi:10.1016/j.acra.2026.07.024
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。