第一部分：基于CT图像直方图和纹理特征分析鉴别长骨骨肉瘤与尤文肉瘤的多中心研究

【背景与目的】骨肉瘤（OS）与尤文肉瘤（ES）是长骨常见的原发恶性骨肿瘤，两者CT影像表现相似，传统影像学检查难以准确鉴别。纹理分析通过提取医学图像中的纹理特征，对病变或组织内像素或体素的灰度分布定量分析，能较为客观地对两种肿瘤鉴别诊断。本研究通过建立多个CT影像组学模型并进行定量分析，比较几种模型的鉴别效能，探讨直方图和纹理分析在鉴别长骨OS与ES的应用价值，旨在寻找一种早期无创的方式对两者进行准确鉴别诊断。【材料与方法】回顾性收集三个中心经手术病理证实的长骨OS及ES患者各25例，以8∶2的比例将所有病例随机分为训练集（OS 21例，ES 19例）和验证集（OS 4例，ES 6例）。在CT图像上手动勾画感兴趣区域（ROI）提取纹理特征参数。利用随机森林（RF）及LASSO算法进行特征筛选，构建逻辑回归（LR）、RF、支持向量机（SVM）和K近邻（KNN）四种影像组学模型，绘制受试者工作特征曲线（ROC）并计算曲线下面积（AUC），评价四个模型的诊断效能。【结果】基于CT图像共提取100个纹理参数，经筛选获得8个特征参数（最大3D直径、第10百分位数、峰度、最大像素强度值、逆差分归一化、灰度水平方差、长行程高灰度强调、低灰度区域强调），构建的四种模型在验证组的AUC值分别为0.92、0.79、0.83、0.73，其中LR、SVM算法训练的模型诊断效能较高，LR模型在验证集的准确率为90%，灵敏度为83%，特异度为100%，AUC为92%；SVM模型在验证集的准确率为80%，灵敏度为67%，特异度为100%，AUC为83%。【结论】基于CT图像建立的LR、SVM模型对于鉴别长骨OS和ES具有较高价值。【关键词】骨肉瘤；尤文肉瘤；计算机体层成像；纹理分析

第二部分：基于增强T₁WI和T₂WI-FS的影像组学分析在长骨骨肉瘤和尤文肉瘤鉴别诊断中的价值：一项多中心研究

【背景和目的】影像学检查是骨肉瘤（OS）与尤文肉瘤（ES）诊断与鉴别诊断的重要方法，磁共振成像（MRI）具有良好的软组织分辨率，是骨肿瘤患者确定肿瘤体积、评估局部浸润程度的首选影像检查方法。T1抑脂增强序列（增强T₁WI）和T2抑脂序列（T₂WI-FS）作为MRI的常用检查序列，为临床诊断及局部分期提供重要信息。然而长骨OS与ES的MRI影像表现相似，仅依靠传统影像技术准确鉴别面临挑战。本研究基于增强T₁WI和T₂WI-FS提取影像组学特征建立模型并进行分析，通过多个指标比较几种影像组学模型的鉴别效能，探讨基于多序列MRI的影像组学鉴别长骨OS与ES的预测效能。【材料与方法】回顾性收集三个中心经手术病理证实的25例长骨OS及24例ES患者，以7∶3的比例将所有病例随机分为训练集（OS 19例，ES 15例）和验证集（OS 6例，ES 9例）。在增强T₁WI和T₂WI-FS图像上手动勾画感兴趣区域（ROI）提取纹理特征参数，使用独立样本t检验、Person相关性分析和5折交叉验证的LASSO方法从原始数据集中筛选重要特征，采用逻辑回归（LR）、极端梯度提升（XGBoost）、K近邻（KNN）、高斯贝叶斯、决策树、随机森林（RF）及支持向量机（SVM）7种分类器构建模型，绘制受试者工作特征曲线（ROC）并计算曲线下面积（AUC），评价7个模型的诊断效能。【结果】从MRI 增强T1WI和T₂WI-FS图像分别提取影像组学特征并进行筛选，基于增强T₁WI、T₂WI-FS和增强T₁WI+T₂WI-FS联合组三个特征集分别构建7种影像组学模型并进行比较分析。在建立模型的测试集中，基于T₂WI-FS和增强T₁WI+T₂WI-FS序列构建的模型较增强T₁WI效能更好，其中LR模型在增强T₁WI、T₂WI-FS和增强T₁WI+T₂WI-FS序列上的AUC值分别达到0.81、0.93和0.93，较其他模型性能更好；基于增强T₁WI+T₂WI-FS联合特征集的LR、KNN及XGBoost模型在测试集的AUC值分别为0.93、0.91、0.85，均达到良好的鉴别水平。【结论】基于多序列MRI的影像组学分析在鉴别长骨OS与ES中有一定潜力，其中LR模型对于两者鉴别有较高的应用价值。【关键词】骨肉瘤；尤文肉瘤；MRI；影像组学

Part I ：Multicenter Study on Differentiating Osteosarcoma from Ewing’s Sarcoma in Long Bones Based on CT Image Histogram and Texture Feature Analysis

[Background and objective] Osteosarcoma (OS) and Ewing’s sarcoma (ES) are common primary malignant bone tumors in long bones, with overlapping CT imaging manifestations that challenge accurate differentiation using conventional radiological methods. Texture analysis, which involves extracting texture features from medical images and performing quantitative analysis of grayscale distribution patterns within lesions or tissues, offers an objective approach for distinguishing these two malignancies. This study aimed to establish and compare multiple CT radiomics models through quantitative analysis, evaluate their diagnostic efficacy, and explore the utility of histogram and texture analysis in differentiating long bone OS from ES, thereby identifying a non-invasive method for early and accurate discrimination.[Materials and Methods] A retrospective cohort of 25 pathologically confirmed long bone OS cases and 25 ES cases from three medical centers was enrolled. The dataset was randomly divided into a training set (21 OS, 19 ES) and a validation set (4 OS, 6 ES) at an 8:2 ratio. Regions of interest (ROI) were manually delineated on CT images to extract texture features. Feature selection was performed using Random Forest (RF) and LASSO algorithms. Four radiomics models—Logistic Regression (LR), RF, Support Vector Machine (SVM), and K-Nearest Neighbors (KNN)—were constructed. Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC) values were generated to evaluate diagnostic performance.[Result] A total of 100 texture parameters was extracted from CT images. After feature selection, 8 parameters were retained: maximum 3D diameter, 10th percentile, kurtosis, maximum pixel intensity, inverse difference normalized, gray-level variance, long-run high gray-level emphasis, and low gray-level zone emphasis. The validation set AUC values for the four models were 0.92 (LR), 0.79 (RF), 0.83 (SVM), and 0.73 (KNN). The LR model demonstrated superior performance in the validation set, with 90% accuracy, 83% sensitivity, 100% specificity, and an AUC of 0.92. The SVM model achieved 80% accuracy, 67% sensitivity, 100% specificity, and an AUC of 0.83.[Conclusion] The LR and SVM models developed from CT-based radiomics exhibit high diagnostic value in differentiating long bone OS from ES.[Keywords] Osteosarcoma; Ewing’s Sarcoma; Computed Tomography; Texture Analysis

Part II Value of Radiomics Analysis Based on CE T₁WI and T₂WI-FS in Differentiating Osteosarcoma from Ewing’s Sarcoma in Long Bones: A Multicenter Study

[Background and objective] Imaging examinations play a critical role in the diagnosis and differentiation of osteosarcoma (OS) and Ewing’s sarcoma (ES). Magnetic resonance imaging (MRI), with its superior soft-tissue resolution, is the preferred modality for determining tumor volume and assessing local infiltration in bone tumor patients. Fat-suppressed contrast-enhanced T₁-weighted imaging (CE T₁WI) and fat-suppressed T₂-weighted imaging (T₂WI-FS), as commonly used MRI sequences, provide essential information for clinical diagnosis and local staging. However, the overlapping MRI manifestations of long bone OS and ES pose challenges for accurate differentiation using conventional imaging techniques. This study aimed to extract radiomic features from CE T₁WI and T₂WI-FS sequences, construct predictive models, and evaluate their diagnostic performance through multiple metrics, thereby exploring the potential of multi-sequence MRI-based radiomics in distinguishing long bone OS from ES.[Materials and Methods] A retrospective cohort of 25 pathologically confirmed long bone OS cases and 24 ES cases from three medical centers was enrolled. The dataset was randomly divided into a training set (19 OS, 15 ES) and a test set (6 OS, 9 ES) at a 7:3 ratio. Regions of interest (ROI) were manually delineated on CE T₁WI and T₂WI-FS images to extract texture features. Feature selection was performed using independent-sample t-tests, Pearson correlation analysis, and LASSO regression with 5-fold cross-validation. Seven classifiers—Logistic Regression (LR), Extreme Gradient Boosting (XGBoost), K-Nearest Neighbors (KNN), Gaussian Naive Bayes, Decision Tree, Random Forest (RF), and Support Vector Machine (SVM)—were employed to construct radiomics models. Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC) values were calculated to evaluate diagnostic performance.[Result] Radiomic features were extracted and filtered from CE T₁WI and T₂WI-FS images. Seven models were constructed and compared based on three feature sets: CE T₁WI alone, T₂WI-FS alone, and a combined CE T₁WI + T₂WI-FS set. In the test set, models built using T₂WI-FS and the combined CE T₁WI +T₂WI-FS feature sets outperformed those based on CE T₁WI alone. The LR model achieved the highest performance, with AUC values of 0.81 (CE T₁WI), 0.93 (T₂WI-FS), and 0.93 (combined set). Among the combined feature set models, LR, KNN, and XGBoost demonstrated strong discriminatory power, with AUC values of 0.93, 0.91, and 0.85, respectively.[Conclusion] Multi-sequence MRI-based radiomics analysis shows promise in differentiating long bone OS from ES, with the LR model exhibiting particularly high diagnostic value.[Key words] Osteosarcoma; Ewing’s Sarcoma; MRI; Radiomics