Prostate biopsy is considered a standard diagnostic tool for prostate cancer (PCa), however it may yield false-negative results or to an over diagnosis of low risk PCa. Improvements of prostate imaging via multi-parametric MRI may lead to better tumor localization and greater accuracy in predicting histopathology and tumor volume of lesions larger than 5 mm (1, 2). Utilizing dynamic contrast-enhanced (DCE)-MRI in imaging PCa helps visualize and characterize the heterogeneity of the tumor. The focus of this study is to apply a promising statistical methodology, k-means clustering, to current state of the art MRI including DCE-imaging and pharmacokinetic parameters based assessment to help classify microcirculatory characteristics of PCa based on voxel level tissue characteristics and correlation with Gleason Score (GS) and pathologic (p) stage.

Subjects: 25 male patients were recruited in the study who were diagnosed with PCa on needle prostate biopsy. Patients underwent 3T multi-parametric prostate MRI (Achieva; Philips Healthcare) with the use of 32-channel phased-array surface coil and parallel radiofrequency transmission (mTX) (3), prior to robot-assisted laparoscopic prostatectomy (RALP). Histopathology staging is also available.

MRI: The DCE-MRI was acquired using 3D spoiled gradient echo, TR/TE: 5/1.5 milliseconds; flip angle: 20 degrees; slice thickness: 6 mm; in-plane resolution: 0.78 × 0.78 -0.90 × 0.90 mm2; temporal resolution: 5.4-14.1 seconds; number of dynamic scans: 30-59. Bolus of Gd-based contrast agent was intravenously injected at a constant flow rate of 0.5 mL/s after the fifth phase.

Image processing and analysis: DCE data were processed using in house software written in interactive data language (IDL). Modified Brix’s linear two compartment model was used to estimate the voxel-wise pharmacokinetic parameters, known as the amplitude of signal enhancement (Amp) and the kep was defined as the exchange rate of the contrast agent between extracellular-extravascular space (EES) and the plasma. Tumor localization and placement of tumor regions of interest (ROI) were reviewed by a radiologist and pathologist on the respective data sets. The K-means clustering method was used to data mine, each cluster has a centroid, the inclusion-exclusion principle for each data point determined by its distance from each cluster centroid as shown in figure 1.

The k-means clustering differentiated tumors into cluster 1 (low kep-low Amp), cluster 2 (low kep-high Amp), and cluster 3 (high kep-low Amp) as demonstrated in figure 2. Tumor volume fractions (VF) of cluster 1, 2 and 3 had mean values of 41±11%, 34±15%, and 25±10%, respectively. Negative correlations of high Amp-p-stage (r=-0.5), high Amp-VF2 (r=-0.7) were found within cluster 2, a negative correlation of high kep-VF3 (r=-0.6) and low Amp-VF3 (r=-0.5) were found within cluster 3. Cluster 3 Amp-kep values were higher in p-stage 3 than p-stage 2.The use of k-means clustering allowed better classification of pharmacokinetic as well as microcirculatory characteristics of PCa in correlation with p-stage and VF. Our results were similar to previously published study performed in bladder cancer patients receiving neoadjuvant chemotherapy (4). We believe, the pixel by pixel analysis of the PCa-ROI identifying each cluster percentage and its perfusion features and combining those with the clinical findings, would assist in therapeutic treatment decision-making introducing an exciting future methodology to utilize functional MRI in managing PCa.