基于多参数MRI影像组学与深度学习预测高危多发性骨髓瘤的研究

Predicting High-Risk Multiple Myeloma Based on Multiparametric MRI Radiomics and Deep Learning

摘要: 目的：本研究旨在构建基于多参数MRI的影像组学模型、深度学习模型及影像组学深度学习联合模型，并评估不同模型预测高危多发性骨髓瘤的效能。方法：本研究回顾性收集了青岛大学附属医院从2014年5月至2025年3月经病理确诊的140例多发性骨髓瘤脊椎MRI图像，共258个病变，按照8:2的比例随机分为训练集(n = 206)、验证集(n = 52)。基于MRI常规序列矢状位T1WI、T2WI和FS-T2WI图像，使用ITK-SNAP软件对最大病灶累及椎体逐层勾画感兴趣区(Region of interest, ROI)，人工提取影像组学特征，并采用2.5D Densenet169深度学习模型自动提取深度学习特征，分别构建影像组学模型和深度学习模型，然后通过多元逻辑回归将选择的影像组学特征和深度学习特征融合构建影像组学深度学习联合模型。采用受试者工作特征曲线(Operating characteristics curve, ROC)、曲线下面积(Area under curve, AUC)、灵敏度、特异度和准确度评估影像组学模型、深度学习模型和影像组学深度学习联合模型的预测效能。采用决策曲线分析(Decision curve analysis, DCA)比较三种模型在不同决策阈值下的净获益。结果：最终选择了4个影像组学特征和1个深度学习特征构建影像组学深度学习联合模型。影像组学深度学习联合模型的AUC值(训练集：0.996；验证集：0.893)高于影像组学模型和深度学习模型。决策曲线分析结果显示影像组学深度学习联合模型在预测高危多发性骨髓瘤方面具有更好的临床适用性。结论：影像组学深度学习联合模型在预测高危多发性骨髓瘤方面相较于其他两种模型具有更好的效能，有助于制定临床治疗策略。

Abstract: Objective: This study aimed to construct radiomics models, deep learning models, and integrated radiomics-deep learning models based on multiparametric MRI, and to evaluate the performance of different models in predicting high-risk multiple myeloma. Methods: This study retrospectively collected 140 patients in the Affiliated Hospital of Qingdao University from May 2014 and March 2025, including 258 vertebral lesions, all patients were pathologically confirmed multiple myeloma. The lesions were randomly divided into training (n = 206) and validation (n = 52) sets at an 8:2 ratio. Based on conventional MRI sequence, including sagittal T1-weighted imaging(T1WI), T2-weighted imaging (T2WI), and fat-suppressedT2-weighted imaging (FS-T2WI). The ROI was manually segmented on the largest lesion-involved vertebral body using ITK-SNAP software. Radiomics features were manually extracted, while a 2.5D DenseNet169 deep learning model was employed to automatically extract deep learning features. Separate radiomics and deep learning models were constructed, followed by an integrated radiomics-deep learning model developed by combining selected features through multivariate logistic regression. The predictive efficiency of the radiomics, deep learning, and integrated models was evaluated using receiver operating characteristic curves (ROC), area under the curve (AUC), sensitivity, specificity, and accuracy. Decision curve analysis (DCA) was applied to compare the net benefits of the three models across different decision thresholds. Results: A combined radiomics-deep learning model was constructed using four radiomic features and one deep learning feature. The AUC values of the combined radiomics-deep learning model (training set: 0.996; validation set: 0.893) were higher than those of the radiomics model and the deep learning model. Decision curve analysis results demonstrated that the combined radiomics-deep learning model has better clinical applicability in predicting high-risk multiple myelomacon. Conclusion: The integrated radiomics-deep learning model demonstrated superior performance in predicting high-risk multiple myeloma compared to the other two models, suggesting its potential for guiding clinical treatment strategies.