【摘要】 目的：探讨基于腹部CT平扫图像提取的脾脏影像组学特征构建的机器学习模型对急性髓系白血病（AML）患者疾病状态的预测效能。方法：回顾性纳入259例AML患者，其中疾病状态（包括初诊、复发、治疗无反应）151例，完全缓解108例；按7∶3比例随机分为训练集181例和测试集78例。将所有患者的腹部CT平扫图像，经过重采样和标准化，使用3D Slicer软件半自动勾画脾脏为VOI。基于PyRadiomics软件包提取几何特征、灰度特征、纹理特征等影像组学特征。使用Z-score对特征数据正则化后依次采用t检验（P<0.05）和Pearson相关分析（r>0.9）行初步降维，随后使用最小绝对收缩和选择算子（LASSO）算法进行特征筛选，保留最优特征，输入多种机器学习算法行10折交叉验证学习，建立相应模型，并在测试集中通过AUC、准确率、敏感度、特异度及F1分数评估模型性能。结果：PyRadiomics软件包共提取1 561个影像组学特征，最终筛选保留最优的17个特征，并建立11种机器学习模型。在测试集中，Naive Bayes模型展现出最佳的总体预测效能，其AUC、准确率、敏感度、特异度和F1分数分别为0.811、0.782、0.780、0.784和0.790;LR模型次之；MLP模型AUC和特异度最高（0.828和0.919），但敏感度欠佳，提示其对疾病状态的预测结果高度可信。结论：基于脾脏CT影像组学特征构建的机器学习模型能有效鉴别AML患者的疾病或完全缓解状态。其中，Naive Bayes、MLP、LR模型在测试集上性能优异，有望为AML患者的疾病状态评估提供一种无创、客观的辅助诊断工具。更多还原

【Abstract】 Objective:To explore the value of machine learning models based on splenic radiomic features from non-contrast abdominal CT scan in the prediction of disease status or post-treatment complete remission(CR) in patients with acute myeloid leukemia(AML). Methods:A retrospective study was conducted on 259 patients with AML. Based on bone marrow cytology findings,the patients were categorized into a disease group(including newly diagnosed,relapsed or non-remission patients,151 cases) and a CR group(108 cases),and randomly divided into a training set(181 cases) and a test set(78 cases) in a 7∶3 ratio. Non-contrast abdominal CT images were resampled and standardized,then the spleen was semi-automatically segmented as VOI using 3D Slicer software. The geometric,intensity and texture features were extracted using PyRadiomics software. After Z-score normalization,preliminary feature reduction was performed using t-test(P<0.05) and Pearson correlation analysis(r>0.9),followed by feature selection via LASSO algorithm to retain the optimal features. These selected features were then input into multiple machine learning algorithms for ten-fold cross-validation training to establish the machine learning models. The model performance was evaluated on the test set using AUC,accuracy,sensitivity,specificity,and F1 score. Results:A total of 1 561 radiomic features were extracted,17 optimal features were retained,and 11 machine learning models were established. In the test set,Naive Bayes model demonstrated the best overall predictive performance,with the AUC of 0.811,accuracy of 0.782,sensitivity of 0.780,specificity of 0.784,and F1 score of 0.790,followed by LR model. MLP model achieved the highest AUC and specificity(0.828 and 0.919),though its sensitivity was relatively lower. Conclusions:Machine learning models constructed from splenic CT radiomic features can effectively discriminate between AML patients in disease status and those in CR. Naive Bayes,MLP,and LR models exhibit excellent performance in the test set,showing promise as non-invasive and objective auxiliary diagnostic tools for assessing AML disease status,with potential for clinical translation.更多还原