机器学习算法预测丘脑中央中核立体定向坐标的比较分析
本研究基于100例健康人MRI开发并优化四种机器学习模型,并在20例全面性癫痫患者中验证,用于预测丘脑中央中核的立体定向坐标。深度神经网络预测精度最高,健康数据集平均欧氏误差为0.88±0.41 mm,癫痫数据集为1.12±0.44 mm。该方法可减少对先进成像技术的依赖,使中央中核脑深部电刺激更可及。
要点
- 深度神经网络基于常规T1加权MRI可较准确预测中央中核坐标。
- 健康数据集平均欧氏误差0.88±0.41 mm,癫痫数据集1.12±0.44 mm。
- 线性回归、支持向量回归和K近邻模型表现相似但误差更高。
核心结论
- 机器学习模型尤其是深度神经网络,能基于标准T1加权MRI准确预测中央中核坐标。
- 该方法可降低对先进成像技术的依赖,使中央中核脑深部电刺激在资源有限环境下更可及。
研究设计与样本
- 研究纳入100例健康人MRI开发并优化四种机器学习模型,另在20例全面性癫痫患者中独立验证。
- 四种模型为线性回归、K近邻、支持向量回归和深度神经网络。
- 模型输入特征为T1加权MRI上易识别点的x、y、z坐标,输出为中央中核立体定向坐标。
主要发现
- 深度神经网络预测中央中核坐标精度最高,健康数据集平均欧氏误差0.88±0.41 mm。
- 深度神经网络在癫痫数据集平均欧氏误差为1.12±0.44 mm。
- 线性回归、支持向量回归和K近邻模型表现相似,但误差率均高于深度神经网络。
临床怎么用
- 可利用常规T1加权MRI和机器学习模型辅助预测中央中核坐标,减少对先进成像技术的依赖。
- 深度神经网络模型在健康人和全面性癫痫患者中均表现出较低预测误差,可能有助于提高电极放置准确性。
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)。
原文
OBJECTIVE: Deep brain stimulation (DBS) of the centromedian nucleus (CM) of the thalamus is a promising treatment for drug-resistant epilepsy, Tourette syndrome, disorders of consciousness, and chronic pain, particularly when other surgical options are not feasible. However, the CM is challenging to visualize on routine MRI and atlas-based targeting often results in inaccurate electrode placement, affecting surgical outcomes. Furthermore, inability to visualize and directly target the CM is a barrier to CM-DBS in a resource-limited setting. The aim of this study was to develop and test machine learning (ML) models that could predict target coordinates of the CM using multiple datapoints available from conventional T1-weighted MRI. METHODS: Four ML models-linear regression (LR), k-nearest neighbor (KNN), support vector regression (SVR), and deep neural network (DNN)-were developed and optimized using 100 MR images obtained in healthy individuals and validated in a separate dataset of 20 patients with generalized epilepsy, which is an indication for CM-DBS. Models were trained to predict the stereotactic coordinates of the CM using input features, which were x, y, and z coordinates of readily identifiable points from T1-weighted MRI. RESULTS: The DNN model demonstrated the highest accuracy in predicting CM coordinates, with a mean Euclidean error of 0.88 ± 0.41 mm in the healthy dataset, and 1.12 ± 0.44 mm in the epilepsy dataset. The LR, SVR, and KNN models all performed similarly, although with higher error rates. CONCLUSIONS: These findings indicate that ML models, particularly DNNs, can accurately predict CM coordinates using standard T1-weighted MRI. This approach reduces dependency on advanced imaging techniques, making CM-DBS more accessible.
引用
- Singh H, Bhatia N, Shless J, et al. Comparative analysis of machine learning algorithms for predicting stereotactic coordinates of the centromedian nucleus. J Neurosurg. 2026. PMID: 41569646. doi:10.3171/2025.8.JNS243201
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。