3.0T磁共振图像计算机特征鉴别诊断前列腺良恶性病变的价值

Computer-aided analysis of 3.0T magnetic resonance image features in differentiation of benign and malignant prostate lesions

关键词:  前列腺疾病  磁共振成像  图像特征  诊断

Key words:  prostate diseases  magnetic resonance imaging  image feature analysis  diagnosis

摘要:

目的 探讨3.0T MRI图像计算机特征鉴别诊断前列腺良恶性病变的价值。方法 收集接受3.0T磁共振检查的前列腺患者330例，进行T1加权成像（T2WI）、T2加权成像（T2WI）、多b值的弥散成像（DWI）与动态增强检查。所有病例均经活检或手术病理证实，其中前列腺癌患者198例，前列腺增生患者132例。选取病变最大层面图像，手动勾画病灶感兴趣区（ROI），进行方向梯度直方图（HOG）特征、局部二值模式（LBP特征）、哈尔特征（Haar特征）提取。通过支持向量机（SVM）分类器进行训练并得到对应的分类模型，通过不同的分类模型对比选取最具有鉴别前列腺良恶性病变价值的图像特征参数，并采用图像特征分类分析统计方法对所选择的图像特征参数进行评估。鉴别诊断结果以绘制受试者工作特征曲线（ROC曲线）表示，并通过计算曲线下面积（AUC）、灵敏度、特异度、准确率来进行效果评估。结果 磁共振T2WI、长b值DWI、ADC图中，在Haar图像特征条件下，ADC图的AUC值最大（0.85）；在HOG图像特征条件下，DWI序列的AUC值最大（0.78）；在LBP图像特征条件下，ADC图的AUC值最大（0.87）；在采用三种特征融合[HOG+LBP+Haar（HLH）]的图像特征条件下，DWI序列的AUC值较大（0.89），大于ADC图及T2WI序列的AUC值（0.88、0.84）。图像特征分类分析方法中，LBP图像特征区分两种病变的AUC值较大（0.87），大于Haar图像特征（0.85）和HOG图像特征（0.78），HLH融合特征方法的AUC值（0.89）大于其中任一特征，具有最优的鉴别诊断结果。结论 3.0T MRI的T2WI、DWI/ADC图像计算机特征分析有助于鉴别诊断前列腺良恶性病变，长b值的DWI和ADC图对鉴别诊断前列腺良恶性结节占有重要地位，而HLH融合的计算机图像特征分析能为前列腺良恶性病变鉴别诊断提供可靠依据。

Abstract:

Objective To investigate the diagnostic value of computer-aided analysis of image features on 3.0T MRI in differentiating benign and malignant prostate lesions. Methods Three hundred and thirty patients with prostate diseases underwent 3.0 Tesla MRI (3.0 Tesla MRI, MagnetomVerio, Siemens ) examination. T1 weighted imaging (T1WI), T2 weighted imaging (T2WI) and multi-b-value diffusion imaging (DWI) and dynamic contrast-enhancement were performed 7d before operation. Among 330 patients, 198 cases of prostatic cancer and 132 cases of prostatic hyperplasia were confirmed by biopsy or surgery pathology. The region of interest (ROI) was delineated manually from the max-level image of the lesion. The HOG feature, local binary pattern (LBP feature) and Haar feature were extracted. The support vector machine (SVM) classifier was trained and the corresponding classification model was obtained. By comparing different classification models, the most valuable image feature parameters for distinguishing benign and malignant prostatic lesions were selected, and the selected image feature parameters for differential diagnosis were evaluated by image feature classification analysis and statistics method. The efficiency of differential diagnosis was evaluated by calculating the area under the ROC curve (AUC), sensitivity, specificity and accuracy. Results For MR images of T2WI, DWI and ADC diagram, the AUC value of the ADC graph was the largest (0.85) under the Haar image feature condition, the AUC value of DWI sequence was the largest (0.78) under the condition of HOG image feature and the AUC value of ADC diagram was the largest (0.87) under the condition of LBP image feature. However, under the image feature conditions of three feature fusion HOG+LBP+Haar (HLH), the AUC value of the DWI sequence (0.89) was larger than that of the ADC map (0.88) and T2WI sequence (0.84). In the method of classification and analysis of image features, the AUC value of LBP image feature (0.87) in differentiating malignant from benign lesions was higher than that of Haar image feature (0.85) and HOG image feature (0.78). The AUC value (0.89) of the HLH fusion feature method was larger than any other features and had the best differential diagnostic effect. Conclusion The computer-aided analysis of T2WI and DWI/ADC image features for 3.0T MRI will be helpful to differentiate benign and malignant prostate lesions, and long b value DWI and ADC image analysis play an important role. HLH computer-aided analysis of image features may provide a reliable basis in identifying benign and malignant prostate lesions