基于病灶内外多模态MRI影像组学特征构建可解释性机器学习模型鉴别布鲁杆菌和结核性脊柱炎

Differentiating brucellar spondylitis from tuberculous spondylitis with an interpretable machine learning model using intra- and peri-lesion radiomics features from multi-sequence MRI

中文关键词:  布鲁杆菌脊柱炎  结核性脊柱炎  多模态影像组学  机器学习  鉴别诊断

英文关键词:Brucellar spondylitis  Tuberculous spondylitis  Multi-modal radiomics  Machine learning  Differential diagnosis

中文摘要:

  【摘要】 目的：基于布鲁杆菌脊柱炎（brucellar spondylitis，BS）与结核性脊柱炎（tuberculous spondylitis，TS）患者病灶内和病灶外多模态影像组学特征，构建可解释性机器学习模型，探讨其在鉴别诊断中的价值。方法：回顾性纳入2020年1月~2024年12月在新疆维吾尔自治区第六人民医院收治的186例BS和TS患者（BS 83例，TS 103例），所有患者均经病理学检查或微生物培养确诊。患者均接受脊柱MRI T1加权成像（T1-weighted imaging，T1WI）、T2加权成像（T2-weighted imaging，T2WI）及脂肪抑制T2加权成像（fat-suppressed T2-weighted imaging，FS-T2WI）序列检查，通过3D Slicer软件勾画病灶感兴趣区域（region of interest，ROI）及周围5mm扩展区域（peri-ROI），分别从各序列中提取1500个影像组学特征，涵盖一阶统计量、纹理特征及形态学特征，构建113种机器学习组合流程，每个流程包含特征筛选与模型拟合两个模块，通过受试者工作特征（receiver operating characteristic，ROC） 曲线下面积（area under the curve，AUC）、准确度、敏感性、特异性及F1分数评估，同时采用校准曲线分析预测概率的可靠性，并利用Shapley加性解释（Shapley additive explanations，SHAP）解析关键特征对模型决策的贡献度。结果：单序列模型T1WI、T2WI及FS-T2WI序列的病灶内特征模型训练集AUC为0.987~0.998，准确度为0.909~0.977。T1WI序列病灶内特征模型的测试集AUC为0.787，假阳性率为0.338；结合病灶周围特征后，模型AUC提升至0.806，准确度增至0.762。T2WI序列病灶内特征模型的测试集AUC为0.782，假阳性率为0.380；结合病灶周围特征后测试集AUC提升至0.824，假阳性率为0.310。FS-T2WI序列的病灶内模型的测试集AUC为0.829，假阳性率为0.296；结合病灶周围特征后测试集AUC提升至0.810，假阳性率为0.296。融合T1WI、T2WI及FS-T2WI三序列的多模态联合模型病灶内外特征，测试集AUC为0.841（95%置信区间：0.777~0.899），准确度0.804，敏感性0.856，特异性0.732，F1分数0.834，假阳性率与假阴性率分别为0.268和0.144。特征分析显示FS-T2WI序列的病灶内小波低-高-低子带（low-high-low subband，LHL）一阶平均绝对差（P=1.56×10-21）及病灶周围灰度依赖矩阵（gray level dependence matrix，GLDM）依赖方差（P=1.68×10-12）为最具鉴别力的特征。SHAP分析发现病灶内特征的贡献度占主导地位（平均SHAP值为0.0709），其中小波LHL平均绝对差（SHAP值为0.0709）和FS序列病灶内小波LHL GLDM依赖熵（SHAP值为0.0588）对分类决策影响最显著。结论：基于多模态MRI影像组学特征的可解释性机器学习模型能够有效捕捉BS与TS的微观结构与病理异质性，可显著提升对两者的鉴别诊断效能。

英文摘要:

  【Abstract】 Objectives: To construct an interpretable machine learning model based on intra- and peri-lesional multi-modal radiomics features of patients with brucellarspondylitis(BS) and tuberculous spondylitis(TS), and to explore the values in differential diagnosis. Methods: A retrospective analysis was conducted on 186 patients with BS(n=83) and TS(n=103) admitted to the Sixth People′s Hospital of Xinjiang Uygur Autonomous Region from January 2020 to December 2024. All patients were confirmed by pathological examination or microbial culture. All patients underwent spinal MRI, including T1-weighted imaging(T1WI), T2-weighted imaging(T2WI), and fat-suppressed T2-weighted imaging(FS-T2WI) sequences. The core lesion area(region of interest, ROI) and a 5mm surrounding expansion area(peri-ROI) were delineatedusing the 3D Slicer software. A total of 1, 500 radiomics features, covering first-order statistics, texture features, and morphological features, were extracted from each sequence. A total of 113 machine learning combination workflows were constructed, each comprising two modules: feature selection and model fitting. The models were evaluated using the receiver operating characteristic(ROC) curve for the area under the curve(AUC), for accuracy, sensitivity, specificity, and F1 score. Meanwhile, calibration curves were used to analyze the reliability of predicted probabilities, and Shapley additive explanations(SHAP) values were utilized to interpret the contribution of key features to model decision-making. Results: The training set AUCs of the intra-lesional feature models for single-sequence T1WI, T2WI, and FS-T2WI ranged from 0.987 to 0.998, with accuracies ranging from 0.909 to 0.977. For the T1WI intra-lesional feature model, the test set AUC was 0.787 with a false positive rate of 0.338; After combining peri-lesional features, the model AUC increased to 0.806, and accuracy increased to 0.762. For the T2WI intra-lesional feature model, the test set AUC was 0.782 with a false positive rate of 0.380; After combining peri-lesional features, the test set AUC increased to 0.824, and the false positive rate decreased to 0.310. For the FS-T2WI intra-lesional model, the test set AUC was 0.829 with a false positive rate of 0.296; After combining peri-lesional features, the test set AUC increased to 0.810, with a false positive rate of 0.296. The multi-modal combined model, fusing intra- and peri-lesional features from T1WI, T2WI, and FS-T2WI sequences, achieved a test set AUC of 0.841(95% confidence interval: 0.777-0.899), an accuracy of 0.804, a sensitivity of 0.856, a specificity of 0.732, F1 score of 0.834, and false positive and false negative rates of 0.268 and 0.144, respectively. Feature analysis revealed that the wavelet low-high-low(LHL) subband first-order mean absolute deviation of the intra-lesional FS-T2WI sequence(P=1.56×10-21) and the dependence variance of the peri-lesional gray level dependence matrix(GLDM)(P=1.68×10-12) were the most discriminative features. SHAP interpretability analysis found that intra-lesional features dominated the contribution(mean SHAP value of 0.0709), among which the wavelet LHL mean absolute deviation(SHAP value=0.0709) and the intra-lesional FS wavelet LHL GLDM dependence entropy(SHAP value=0.0588) had the most significant impact on classification decisions. Conclusions: The interpretable machine learning model based on multi-modal MRI radiomics features can effectively capture the microstructural and pathological heterogeneity between BS and TS, significantly improving the differential diagnostic efficacy for both diseases.