DataSheet_1_Assessing and testing anomaly detection for finding prostate cancer in spatially registered multi-parametric MRI.pdf

BackgroundEvaluating and displaying prostate cancer through non-invasive imagery such as Multi-Parametric MRI (MP-MRI) bolsters management of patients. Recent research quantitatively applied supervised target algorithms using vectoral tumor signatures to spatially registered T1, T2, Diffusion, and Dynamic Contrast Enhancement images. This is the first study to apply the Reed-Xiaoli (RX) multi-spectral anomaly detector (unsupervised target detector) to prostate cancer, which searches for voxels that depart from the background normal tissue, and detects aberrant voxels, presumably tumors.
MethodsMP-MRI (T1, T2, diffusion, dynamic contrast-enhanced images, or seven components) were prospectively collected from 26 patients and then resized, translated, and stitched to form spatially registered multi-parametric cubes. The covariance matrix (CM) and mean μ were computed from background normal tissue. For RX, noise was reduced for the CM by filtering out principal components (PC), regularization, and elliptical envelope minimization. The RX images were compared to images derived from the threshold Adaptive Cosine Estimator (ACE) and quantitative color analysis. Receiver Operator Characteristic (ROC) curves were used for RX and reference images. To quantitatively assess algorithm performance, the Area Under the Curve (AUC) and the Youden Index (YI) points for the ROC curves were computed.
ResultsThe patient average for the AUC and [YI] from ROC curves for RX from filtering 3 and 4 PC was 0.734[0.706] and 0.727[0.703], respectively, relative to the ACE images. The AUC[YI] for RX from modified Regularization was 0.638[0.639], Regularization 0.716[0.690], elliptical envelope minimization 0.544[0.597], and unprocessed CM 0.581[0.608] using the ACE images as Reference Image. The AUC[YI] for RX from filtering 3 and 4 PC was 0.742[0.711] and 0.740[0.708], respectively, relative to the quantitative color images. The AUC[YI] for RX from modified Regularization was 0.643[0.648], Regularization 0.722[0.695], elliptical envelope minimization 0.508[0.605], and unprocessed CM 0.569[0.615] using the color images as Reference Image. All standard errors were less than 0.020.
ConclusionsThis first study of spatially registered MP-MRI applied anomaly detection using RX, an unsupervised target detection algorithm for prostate cancer. For RX, filtering out PC and applying Regularization achieved higher AUC and YI using ACE and color images as references than unprocessed CM, modified Regularization, and elliptical envelope minimization.

背景 通过多参数磁共振成像（MP-MRI）等非侵入性影像手段评估与展示前列腺癌，可有效优化患者诊疗管理。既往研究已将结合向量肿瘤特征的监督目标算法，定量应用于空间配准的T1加权、T2加权、弥散加权及动态对比增强成像。本研究首次将Reed-Xiaoli（RX）多光谱异常检测器（无监督目标检测器）应用于前列腺癌检测，该检测器通过搜寻偏离背景正常组织的体素，识别疑似肿瘤的异常体素。
方法 前瞻性收集26例患者的多参数磁共振成像（MP-MRI）图像（包含T1加权、T2加权、弥散加权、动态对比增强成像共7个分量），随后对图像进行尺寸调整、配准与拼接，构建空间配准的多参数体数据块。从背景正常组织中计算协方差矩阵（CM）与均值μ。针对RX算法，通过剔除主成分（PC）、正则化及椭圆包络最小化来降低协方差矩阵的噪声。将RX算法生成的图像与基于阈值自适应余弦估计器（ACE）及定量颜色分析得到的图像进行对比。采用受试者工作特征（ROC）曲线评估RX算法与参考图像的性能，并计算ROC曲线的曲线下面积（AUC）与尤登指数（YI）以量化算法表现。
结果 以自适应余弦估计器（ACE）图像作为参考时，剔除3个与4个主成分的RX算法的曲线下面积（AUC）与尤登指数（YI）平均值分别为0.734[0.706]与0.727[0.703]；改进型正则化RX算法的AUC[YI]为0.638[0.639]，普通正则化RX算法为0.716[0.690]，椭圆包络最小化RX算法为0.544[0.597]，未处理协方差矩阵的RX算法为0.581[0.608]。以定量颜色图像作为参考时，剔除3个与4个主成分的RX算法的AUC[YI]平均值分别为0.742[0.711]与0.740[0.708]；改进型正则化RX算法的AUC[YI]为0.643[0.648]，普通正则化RX算法为0.722[0.695]，椭圆包络最小化RX算法为0.508[0.605]，未处理协方差矩阵的RX算法为0.569[0.615]。所有标准误均小于0.020。
结论 本研究首次将基于RX无监督目标检测算法的异常检测方法应用于空间配准的多参数磁共振成像（MP-MRI）以检测前列腺癌。相较于未处理协方差矩阵、改进型正则化及椭圆包络最小化方法，剔除主成分并应用正则化的RX算法在以ACE图像与颜色图像作为参考时，可获得更高的曲线下面积与尤登指数。