MRI影像组学特征联合临床特征鉴别布鲁杆菌性脊柱炎和化脓性脊柱炎

A preliminary study of MRI-based radiomics combined with clinical features for Differential Diagnosis of Brucella Spon-dylitis and Pyogenic Spondylitis

目的探讨临床特征及MRI T2加权脂肪抑制像(T2WI-FS)影像组学特征在布鲁杆菌性脊柱炎与化脓性脊柱炎鉴别诊断中的应用价值。方法收集2019年1月至2021年12月新疆医科大学附属第一医院经病理或病原学培养确诊的26例布鲁杆菌性脊柱炎和23例化脓性脊柱炎患者的临床资料。对人口学特征、临床特征及实验室检查等行单因素分析,筛选出有统计学意义的潜在临床危险因素。通过手动勾画术前矢状面T2WI-FS的感兴趣区,利用Pyradiomics包提取多样化的影像组学特征,包括形状、纹理和灰度值等;对影像组学特征进行归一化、冗余性分析排除高度相关的特征,再通过统计方法筛选与研究目标相关的特征;采用最小绝对收缩和选择算子(least absolute shrinkage and selection operator,LASSO)回归,从高维特征中筛选出最具鉴别意义的特征以优化模型的预测性能;利用选出的影像组学特征计算影像组学评分。将筛选的临床危险因素、影像组学特征及影像组学评分纳入logistic回归,构建临床特征模型、影像组学评分模型及临床特征-影像组学评分模型。采用混淆矩阵及受试者操作特征曲线(receiver operating characteristic,ROC)评估模型的鉴别能力。结果49例患者纳入研究,男36例、女13例,年龄(53.79±13.79)岁(范围23~83岁)。化脓性脊柱炎患者的C反应蛋白(C-reaction protein,CRP)及红细胞沉降率(erythrocyte sedimentation rate,ESR)水平高于布鲁杆菌性脊柱炎患者,CRP及ESR是潜在临床危险因素(P<0.05)。共获得影像组学特征1500个,筛选出7个影像组学特征(logarithm glrlm SRLGLE、exponential glcm Imc1、exponential glcm MCC、exponential gldm SDLGLE、square glcm ClusterShade、squareroot glszm SALGLE和wavelet.HHH glrlm Run Variance)。7个影像组学特征中square glcm ClusterShade的效能最好,曲线下面积(area under the curve,AUC)值为0.780、灵敏度68.8%、特异度94.4%、准确度82.4%、精确度91.7%、阴性预测值0.773、阳性预测值0.917;logarithm glrlm SRLGLE的AUC为0.736、灵敏度81.2%、特异度72.2%、准确度76.5%、精确度72.2%、阴性预测值0.812、阳性预测值0.722;exponential glcm Imc1的AUC为0.736、灵敏度50.0%、特异度94.4%、准确度73.5%、精确度88.9%、阴性预测值0.680、阳性预测值0.889。临床特征模型、影像组学评分模型及临床特征-影像组学评分模型的AUC分别为0.801、0.818和0.875。其中临床特征-影像组学评分模型预测效果最好,灵敏度87.5%、特异度77.8%、准确度82.4%、精确度77.8%、阴性预测值0.875。结论临床特征-影像组学评分模型鉴别诊断布鲁杆菌性脊柱炎和化脓性脊柱炎的可行性较高,有助于临床鉴别诊断并提供个体化治疗。

Objective To elucidate the diagnostic utility of clinical features and radiomics characteristics derived from magnetic resonance imaging T2-weighted fat-suppressed images(T2WI-FS)in differentiating brucellosis spondylitis from pyogenic spondylitis.Methods Clinical records of 26 patients diagnosed with Brucellosis Spondylitis and 23 with Pyogenic Spondylitis were retrospectively reviewed from Xinjiang Medical University First Affiliated Hospital between January 2019 and December 2021.Confirmatory diagnosis was ascertained through histopathological examination and/or microbial culture.Demographic characteristics,symptoms,clinical manifestations,and hematological tests were collected,followed by a univariate analysis to discern clinically significant risk factors.For the radiomics evaluation,preoperative sagittal T2WI-FS images were utilized.Regions of interest(ROIs)were manually outlined by two adept radiologists.Employing the Py Radiomics toolkit,an extensive array of radiomics features encompassing shape,texture,and graylevel attributes were extracted,yielding a total of 1,500 radiomics parameters.Fea ture normalzation and redundancy elimination were implemented to optimize the predictive efficacy of the model.Discriminatory radiomics features were identified through statistical methods like t-tests or rank-sum tests,followed by refinement via least abso lute shrinkage and selection operator(LASSO)regression.An integrative logistic regression model incorporated selected clinical risk factors,radiomics attributes,and a composite radiomics score(Rad-Score).The diagnostic performance of three models clini cal risk factors alone,Rad-Score alone,and a synergistic combination were appraised using a confusion matrix and receiver operat ing characteristic(ROC)analysis.Results The cohort comprised 49 patients,including 36 males and 13 females,with a mean age of 53.79±13.79 years.C-reactive protein(CRP)and erythrocyte sedimentation rate(ESR)emerged as significant clinical risk factors(P<0.005).A total of seven discriminative radiomics features(logarithm gIrlm SRLGLE,exponential glcm Imc I,exponen tial glem MCC,exponential gldm SDLCLE,square glcm ClusterShade,squareroot glszm SALGLE and wavelet.HHH glrlm Run V ariance)were isolated through LASSO regression.Among these selected features,the square glemClusterShade feature exhibited the best performance,with an area under the curve(AUC)value of 0.780.It demonstrated a sensitivity of 68.8%,specificity of 94.4%,accuracy of 82.4%,precision of 91.7%,and negative predictive value of 0.773.Furthermore,the logarithm glrlm SRLGLE feature had an AUC of 0.736,sensitivity of 68.8%,specificity of 72.2%,accuracy of 76.5%,precision of 72.2%,and negative pre dictive value of 0.812.The exponential glcm Imc 1 feature had an AUC of 0.736,sensitivity of 50.0%,specificity of 94.4%,accuracy of 73.5%,precision of 88.9%,and negative predictive value of 0.680.Three diagnostic models were constructed:the clinical risk factors model,the radiomics score model,and the integrated model(clinical risk factors+radiomics score),which showed AUC values of 0.801,0.818,and 0.875,respectively.Notably,the integrated model exhibited superior diagnostic efficacy.Conclusion The amalgamation of clinical and radiomics variables within a sophisticated,integrated model demonstrates promising effi-cacy in accurately discriminating between Brucellosis Spondylitis and Pyogenic Spondylitis.This cuttingedge methodology underscores its potential in facilitating nuanced clinical decision-making,precise diagnostic differentiation,and the tailoring of thera-peutic regimens.