背景:目的是评估基于高b值扩散加权成像(DWI)的影像组学特征对膀胱癌分级的可行性,并比较高b值DWI与标准b值DWI的可能优势。
方法:本研究纳入了74例膀胱癌患者。使用3TMRI获取DWI序列,b值为1000、1700和3000s/mm2,并生成相应的ADC图,其次是特征提取。患者被随机分为训练和测试队列,比例为8:2。通过使用Wilcox分析,比较了从ADC1000,ADC1700和ADC3000图获得的影像组学特征,并且仅选择具有显着差异的影像组学特征。对特征选择和建立影像组学模型进行了最小绝对收缩和选择算子方法和逻辑回归。进行受试者操作特征(ROC)分析以评估影像组学模型的诊断性能。
结果:在培训队列中,ADC1000,ADC1700和ADC3000模型用于区分低级别和高级别膀胱癌的AUC分别为0.901,0.920和0.901.在测试队列中,ADC1000,ADC1700和ADC3000的AUC分别为0.582,0.745和0.745.
结论:从ADC1700图谱中提取的影像组学特征比从常规ADC1000图谱中提取的影像组学特征能提高诊断准确性。
BACKGROUND: The aim was to evaluate the feasibility of radiomics features based on diffusion-weighted imaging (DWI) at high b-values for grading bladder cancer and to compare the possible advantages of high-b-value DWI over the standard b-value DWI.
METHODS: Seventy-four participants with bladder cancer were included in this study. DWI sequences using a 3 T MRI with b-values of 1000, 1700, and 3000 s/mm2 were acquired, and the corresponding ADC maps were generated, followed with feature extraction. Patients were randomly divided into training and testing cohorts with a ratio of 8:2. The radiomics features acquired from the ADC1000, ADC1700, and ADC3000 maps were compared between low- and high-grade bladder cancers by using the Wilcox analysis, and only the radiomics features with significant differences were selected. The least absolute shrinkage and selection operator method and a logistic regression were performed for the feature selection and establishing the radiomics model. A receiver operating characteristic (ROC) analysis was conducted to assess the diagnostic performance of the radiomics models.
RESULTS: In the training cohorts, the AUCs of the ADC1000, ADC1700, and ADC3000 model for discriminating between low- from high-grade bladder cancer were 0.901, 0.920, and 0.901, respectively. In the testing cohorts, the AUCs of ADC1000, ADC1700, and ADC3000 were 0.582, 0.745, and 0.745, respectively.
CONCLUSIONS: The radiomics features extracted from the ADC1700 maps could improve the diagnostic accuracy over those extracted from the conventional ADC1000 maps.