An imaging-based approach predicts clinical outcomes in prostate cancer through a novel support vector machine classification

Abstract

Preoperatively predict the probability of Prostate cancer (PCa) biochemical recurrence (BCR) is of definite clinical relevance. The purpose of this study was to develop an imaging-based approach in the prediction of 3-years BCR through a novel support vector machine (SVM) classification. We collected clinicopathologic and MR imaging datasets in 205 patients pathologically confirmed PCa after radical prostatectomy. Univariable and multivariable analyses were used to assess the association between MR findings and 3-years BCR, and modeled the imaging variables and follow-up data to predict 3-year PCa BCR using SVM analysis. The performance of SVM was compared with conventional Logistic regression (LR) and D'Amico risk stratification scheme by area under the receiver operating characteristic curve (Az) analysis. We found that SVM had significantly higher Az (0.959 vs. 0.886; p = 0.007), sensitivity (93.3% vs. 83.3%; p = 0.025), specificity (91.7% vs. 77.2%; p = 0.009) and accuracy (92.2% vs. 79.0%; p = 0.006) than LR analysis. Performance of popularized D'Amico scheme was effectively improved by adding MRI-derived variables (Az: 0.970 vs. 0.859, p < 0.001; sensitivity: 91.7% vs. 86.7%, p = 0.031; specificity: 94.5% vs. 78.6%, p = 0.001; and accuracy: 93.7% vs. 81.0%, p = 0.007). Additionally, beside pathological Gleason score (hazard ratio [HR] = 1.560, p = 0.008), surgical-T3b (HR = 4.525, p < 0.001) and positive surgical margin (HR = 1.314, p = 0.007), apparent diffusion coefficient (HR = 0.149, p = 0.035) was the only independent imaging predictor of time to PSA failure. Therefore, We concluded that imaging-based approach using SVM was superior to LR analysis in predicting PCa outcome. Adding MR variables improved the performance of D'Amico scheme.

Keywords: MRI; biochemical recurrence; prostate cancer; radical prostatectomy; support vector machine.

Figure 1. MR images of a representative case to indicate the imaging registration, lesion identification and region of interesting drawing

A 69-year-old man presented with two solid tumor foci in right peripheral zone (PZ; write arrow) (pathological Gleason score 4+4, pT3a) and left PZ (pathological Gleason score 4+3), respectively. The two lesions were characterized with hypo-signal intensity (SI) on T2-weighted imaging (T2WI; a), and hyper-SI on diffusion-weighted imaging (DWI; b = 1000 s/mm²). Registration and fusion of images between T2WI and DWI was performed using a. DICOM-tag metrics c., showing clearly the lesion boundaries. Because the lesion in right TZ had larger size, and an infiltration of the peri-prostatic fat (extracapsular extension) was suspected, it was selected as leading lesion for further quantitative measure.

Figure 2. A multi-parametric prostate MRI in a 68-year-old man (PSA of 34.3 ng/ml, biopsy Gleason score 4+3 and stage T3b) to show the metrics for imaging interpretation

a. Tumor featured with decreased SI on axial T₂WI, the location was defined at PZ (white line). b. ECE (red arrow) and seminal vesicle invasion (SVI; yellow arrow) were notified on coronal T₂WI with fat suppression. c. Tumor with SVI was confirmed on sagittal T₂WI (yellow arrow). d. Tumor was characteristic with hyper-SI on DWI (b = 1000 s/mm²) and decreased ADC e. f. Schematic diagram shows three-type DCE curves, for this patient, tumor was defined as type-3 DCE curve (blue color). Histopathologic results showed a pathological Gleason core 4+4 and surgical SVI in this patient, and PSA failure was determined on 16 months after the prostatectomy treatment.

Figure 3. the comparison of ROC curves among four risk predictive models constructed with different classification methods and input variables

a. with the same MR input variables, the model constructed by support vector machine (SVM_MR) has significantly higher area under the ROC curve value (Az = 0.959) than the model of logistic regression (Az = 0.886, p = 0.007). b. using the same SVM analysis, the model combining MR and DA'mico variables has significantly higher Az (0.970) than the model using sole DA'mico variables (Az = 0.859; p < 0.001).

Figure 4. Predicted BCR-free Kaplan-Meier curves of 205 patients after radical prostatectomy by four constructed models

The curves are stratified by: patients' true Kaplan-Meier curve (black), and predicted Kaplan-Meier curves by LR_MR (green), SVM_MR (blue), SVM_D'Amico (brown) and SVM_D'Amico+MR model (red), respectively. It shows that the Kaplan-Meier functions constructed by SVM_MR and SVM_D'Amico+MR have relatively smaller bias than these of LR_MR and SVM_D'Amico.