Bao Zixu, Zhang Haosu, Schwendner Maximilian, Schröder Axel, Meyer Bernhard, Krieg Sandro M, Ille Sebastian
Department of Neurosurgery, Klinikum Rechts Der Isar, School of Medicine, Technical University of Munich, Munich, Germany.
Department of Neurosurgery, Heidelberg University Hospital, Heidelberg, Germany.
Hum Brain Mapp. 2025 Sep;46(13):e70344. doi: 10.1002/hbm.70344.
Postoperative aphasia (POA) is a common complication in patients undergoing surgery for language-eloquent lesions. This study aimed to enhance the prediction of POA by leveraging preoperative navigated transcranial magnetic stimulation (nTMS) language mapping and diffusion tensor imaging (DTI)-based tractography, incorporating deep learning (DL) algorithms. One hundred patients with left-hemispheric lesions were retrospectively enrolled (43 developed postoperative aphasia, as the POA group; 57 did not, as the non-aphasia (NA) group). Fiber tracking was performed at fractional anisotropy threshold (FAthres) of 0.10 and 0.15, analyzing the total fiber volume (Vfiber), average fractional anisotropy (FA), and number of visualized tracts. Additionally, the visualization ratio (VR) and FA-sensitive visualization were assessed for individual tractography. The NA group demonstrated significantly higher Vfiber (FAthres = 0.10: 61.1 vs. 51.7 cm, p = 0.029; FAthres = 0.15: 36.9 vs. 29.6 cm, p = 0.008), higher FA (FAthres = 0.10: 0.38 vs. 0.35, p = 0.006; FAthres = 0.15: 0.42 vs. 0.39, p = 0.006), and greater tract numbers (FAthres = 0.10: 6.1 vs. 5.7, p = 0.111; FAthres = 0.15: 5.6 vs. 4.8, p = 0.004). Among individual fiber tracts, the corticothalamic fibers (CtF) showed significantly higher VR in the NA group (86.0% vs. 58.1%, p = 0.003), whereas FA-sensitive visualization of CtF was higher in the POA group (11.6% vs. 0.0%, p = 0.013). A binary DL model developed to predict POA achieved a sensitivity of 72.3% and specificity of 85.3%, with an area underthecurve (AUC) of 0.82. Our findings demonstrate the potential of nTMS-based tractography to predict POA by integrating DL. The CtF showed the most significant potential in predicting aphasia risk and understanding the complexity of the language network, whereas their individual predictive contribution within the model remained limited.
术后失语症（POA）是接受语言功能区病变手术患者的常见并发症。本研究旨在通过利用术前导航经颅磁刺激（nTMS）语言图谱和基于扩散张量成像（DTI）的纤维束成像，并结合深度学习（DL）算法，提高对POA的预测能力。回顾性纳入100例左侧半球病变患者（43例发生术后失语症，作为POA组；57例未发生，作为非失语症（NA）组）。在分数各向异性阈值（FAthres）为0.10和0.15时进行纤维束追踪，分析纤维总体积（Vfiber）、平均分数各向异性（FA）和可视化纤维束数量。此外，对个体纤维束成像评估可视化率（VR）和FA敏感性可视化。NA组显示Vfiber显著更高（FAthres = 0.10：61.1对51.7 cm，p = 0.029；FAthres = 0.15：36.9对29.6 cm，p = 0.008），FA更高（FAthres = 0.10：0.38对0.35，p = 0.006；FAthres = 0.15：0.42对0.39，p = 0.006），以及纤维束数量更多（FAthres = 0.10：6.1对5.7，p = 0.111；FAthres = 0.15：5.6对4.8，p = 0.004）。在个体纤维束中，皮质丘脑纤维（CtF）在NA组显示出显著更高的VR（86.0%对58.1%，p = 0.003），而CtF的FA敏感性可视化在POA组更高（11.6%对0.0%，p = 0.013）。开发的用于预测POA的二元DL模型实现了72.3%的敏感性和85.3%的特异性，曲线下面积（AUC）为0.82。我们的研究结果表明基于nTMS的纤维束成像通过整合DL预测POA的潜力。CtF在预测失语症风险和理解语言网络复杂性方面显示出最显著的潜力，但其在模型中的个体预测贡献仍然有限。
