基于椎旁肌肉的自动影像组学模型对腰椎退行性滑脱术后疗效预测*

Prediction of postoperative outcome of lumbar degenerative spondylolisthesis by an automated radiomics model based on paravertebral muscles

摘要:

目的:本研究旨在开发一种基于深度学习的椎旁肌肉分割方法，该方法既高效又具备良好的泛化能力，并使用分割的椎旁肌肉构建的影像组学模型评估该模型在预测腰椎退行性滑脱（lumbar degenerative spondylolisthesis, LDS）患者术后疗效中的应用价值。方法:本研究纳入2021年1月到2023年10月在同济大学附属上海东方医院接受单节段腰椎后路椎间融合术（posterior lumbar interbody fusion, PLIF) 155例LDS病人。按8:2比例分为训练集（124例）和测试集（31例），通过nnU-Net模型进行自动分割轴位T2加权腰椎MRI中腰椎旁肌肉，并提取其影像组学特征，采用T-test、皮尔逊相关性分析和最小绝对收缩和选择算子 (least absolute shrinkage and selection operator, LASSO) 筛选出最具代表性的特征，构建了6种机器学习模型。随后整合临床数据与影像组学特征，建立综合模型。这些模型的性能通过ROC、灵敏度和特异性进行评估，其临床实用性则通过AUC和决策曲线分析（Decision Curve Analysis，DCA）进行评估。结果:XGBoost模型显示出强大的预测性能。整合临床特征与影像组学特征的综合模型在训练队列中的AUC为0.805 (95% CI:0.734-0.87)，在测试队列中AUC达到0.833 (95% CI:0.773-0.894)，显示出显著的预测能力。此外，相较于单独使用的影像组学模型，该综合模型展现出较好的临床实用性，分类准确率也显著提升。结论:研究结果表明，综合模型对LDS患者的术后结果具有较好的预测价值，可作为一种客观的辅助评估工具，为临床医师制定个性化治疗方案提供参考。

Abstract:

Objective:This study aims to develop a deep learning-based method for segmenting paraspinal muscles that is both efficient and possesses strong generalisation capabilities. A radiomics model constructed using the segmented paraspinal muscles is employed to evaluate the model"s applicability in predicting postoperative outcomes for patients with lumbar degenerative spondylolisthesis (LDS). Methods: This retrospective study enrolled 155 patients with LDS who underwent single-segment posterior lumbar interbody fusion (PLIF) at East Hospital affiliated to Tongji University between January 2021 and October 2023. The patients were divided into training set (124 cases) and a test set (31 cases) in an 8:2 ratio, and the lumbar paraspinal muscles were automatically segmented by the nnU-Net model in axial T2-weighted lumbar spine MRI and their radiomics features were extracted with T-test, Pearson"s correlation analysis, and the least absolute shrinkage and selection operator (LASSO) to screen out the most representative features and construct six machine learning models. Subsequently, clinical data and radiomics features were integrated to build comprehensive models. The performance of these models was evaluated by ROC, sensitivity and specificity, and their clinical utility was assessed by AUC and decision curve analysis (DCA). Results: The XGBoost model demonstrated robust predictive performance compared to the other machine learning models evaluated in the study. The combined model, integrating both clinical and radiomic features, exhibited an AUC of 0.805(95% CI:0.734-0.87) in the training set and 0.833 (95% CI:0.773-0.894) in the test set, indicating substantial predictive capability. Moreover, the combined model showed superior clinical benefit and increased classification accuracy when compared to the radiomics model alone.Conclusion: The research findings indicate that the combined model shows good predictive value for postoperative outcomes in LDS patients, serving as an objective auxiliary evaluation tool to provide additional reference for clinicians in formulating personalised treatment plans.