Automatic Detection and Quantitative DCE-MRI Scoring of Prostate Cancer Aggressiveness

Nestor Andres Parra, Alan Pollack, Felix M. Chinea, Matthew C. Abramowitz, Brian Marples, Felipe Munera, Rosa Castillo, Oleksandr N. Kryvenko, Sanoj Punnen, Radka Stoyanova

To develop a robust and clinically applicable automated method for analyzing Dynamic Contrast Enhanced (DCE-) MRI of the prostate as a guide for targeted biopsies and treatments. An unsupervised pattern recognition (PR) method was used to analyze prostate DCE-MRI from 71 sequential radiotherapy patients. Identified regions of interest (ROIs) with increased perfusion were assigned either to the peripheral (PZ) or transition zone (TZ). Six quantitative features, associated with the washin and washout part of the weighted average DCE curve from the ROI, were calculated. The associations between the assigned DCE-scores and Gleason Score (GS) were investigated. A heatmap of tumor aggressiveness covering the entire prostate was generated and validated with histopathology from MRI-ultrasound fused (MRI-US) targeted biopsies. The volumes of the PR-identified ROI's were significantly correlated with the highest GS from the biopsy session for each patient. Following normalization (and only after normalization) with gluteus maximus muscle's DCE signal, the quantitative features in PZ were significantly correlated with GS. These correlations straightened in subset of patients with available MRI-US biopsies when GS from the individual biopsies were used. Area under the receiver operating characteristics curve for discrimination between indolent vs aggressive cancer for the significant quantitative features reached 0.88-0.95. When DCE-scores were calculated in normal appearing tissues, the features were highly discriminative for cancer vs no cancer both in PZ and TZ. The generated heatmap of tumor aggressiveness coincided with the location and GS of the MRI-US biopsies. A quantitative approach for DCE-MRI analysis was developed. The resultant map of aggressiveness correlated well with tumor location and GS and is applicable for integration in radiotherapy/radiology imaging software for clinical translation.