Thonzonium Bromide br Research framework br The research fra
The research framework is summarized in Figure 1. Single parameters were extracted from 43 3-D DCE-US videos of the prostate, and in total, 509 biopsy cores were matched to specific regions in the 3-D parametric maps to evaluate their value as biomarker of PCa. Subse-quently, machine learning by means of an SVM and a GMM was examined for the prediction of SBx out-comes. Its performance was cross-validated in a leave-one-patient-out fashion.
An example of all 3-D parametric maps in a pros-tate harboring sPCa is depicted in Figure 2. The inter-parameter correlations are shown in Figure 3, revealing strong positive and negative correlations in bright green and blue, respectively. In particular, parameters extracted using the same algorithm tended to be more correlated. As for the correlation with the presence of sPCa, in total, 85 parameters showed significant differ-ences in distribution and thus potential as a biomarkers.
As the mean value was generally the most meaningful feature (with 14 of the 16 mean parameter distributions being significant biomarkers), this Thonzonium Bromide measure was more elabo-rately examined. In Figure 4, the per-region-mean parame-ter distributions are depicted using a boxplot with each region displayed as gray dots. The number of regions of each class are shown above the boxplot. Significant or highly significant differences are connected by a line, either marked by * or ** for p < 0.05 and p < 0.005, respectively.
Fig. 1. Schematic overview of model-based machine-learning strategy, consecutively showing (1) 3-D DCE-US data acquisition, (2) model-based feature extraction, (3) region-wise feature extraction and (4) both discriminative and gener-ative machine-learning with ROC-AUC-based feature selection to be validated in a leave-one-patient-out fashion; (5) SBx cores (n = 509) served as ground truth. BPH = benign prostatic hyperplasia; DCE-US = dynamic contrast-enhanced ultrasound; PCa = prostate cancer; ROC-AUC = area under the receiver operating characteristic curve; SBx = systematic biopsy.
ARTICLE IN PRESS
Fig. 2. Example of all 16 3-D parametric prostate maps in the same patient with PCa; the biopsy revealed significant (Gleason 4 + 5) prostate cancer in all three left lateral SBx cores. PCa = prostate cancer; SBx = systematic biopsy; TIC = time-intensity curve.
To obtain the level of significance, a Wilcoxon rank sum test was employed.
In terms of the ROC-AUC, the diagnostic potential to distinguish between either benign tissue (i.e., B, BPH or P) and PCa or benign tissue and sPCa is interstitial listed in Table 1 for all of the single parameters. The wash-in time (WIT) and the convective velocity were the best performing single parameters.
The results of the multi-parametric approach are also shown in Table 2, alongside each classifier’s perfor-mance using only a single parameter. For comparison, the two best performing single parameters for PCa and sPCa were chosen, WIT and vCD, respectively. The GMM especially showed a superior performance com-pared with the single parameters. As the feature selection
ARTICLE IN PRESS
3-D CEUS for Prediction of Prostate Cancer R. R. WILDEBOER et al. 7
Fig. 3. Color-coded inter-parameter correlation matrix of all model-based features, grouped per model parameter; Pear-son correlation coefficients with p > 0.05 are not considered to significantly reflect linear correlation and are marked with a . The two right columns show the Wilcoxon rank sum test’s p value for the association of the features with the presence of PCa or sPCa. a = area under the time-intensity curve; AT = appearance time; DCD = convective dispersion; D = dispersion; FD = fractal dimension; k = dispersion parameter; MI = mutual information; PCa = prostate cancer; Pe = Peclet number; PI = peak intensity; r = spatiotemporal correlation; sPCa = significant prostate cancer; m = mean transit time; vCD = convective velocity; V = velocity ; WIR = wash-in rate; WIT = wash-in time; % = spectral coherence.