Synopsis

Keywords: Prostate, Cancer

This study’s objective was to determine whether automated mpMRI cancer risk maps could identify prostate cancer progression during active surveillance. Derived lesion masks were used to analyze factors for progression. A decision tree model for progression was generated with sensitivity = 0.84, specificity = 0.56, and ROC AUC 0.75. The identification results indicate the potential of mpMRI and MRI cancer risk maps to assist in identifying progression during prostate cancer active surveillance.

Introduction

While prostate cancer (PCa) has high total mortality, many men with PCa may avoid treatment with its concomitant side effects. Improvements in risk assessment are needed to accurately differentiate cases with risk of progression from indolent disease. While multiparametric MRI (mpMRI) is beneficial and widespread, clinical practice relies on qualitative assessments that do not exploit all available information. Applying quantitative methods may aid in detecting and quantifying PCa. This study’s objective was to determine whether mpMRI cancer risk maps could differentiate between men who progressed on active surveillance (AS) versus those who did not.

Methods

This was a retrospective study of PCa progression (defined by an increase in biopsy Gleason grading (GG)) in men following AS. Patients were selected who did not have progression for at least 3 years or had progression in ≤ 3 years from an initial biopsy demonstrating low risk disease, GG = 1. In this cohort, progressed cases had GG = 2 at the later biopsy. Patients were included who had two MRI scans: a) scan 1 near in time to the initial biopsy and b) scan 2 near in time to the progression date or after three years non-progression. Using these criteria and including only those with complete scan sets resulted in 93 subjects (32 progression, 61 non-progression). Lesion masks automatically derived from mpMRI PCa risk maps were used to determine progression indicators. The cancer risk maps were generated according to a methodology previously validated with whole-mount prostate histopathology in men with untreated PCa^1,2,3. The PCa maps were used to classify MRI voxels as cancer from which lesion volumes and mpMRI mean characteristics within lesions were calculated. Investigated parameters included PCa volume, aggressive PCa volume, volumes normalized to the prostate volume, dynamic contrast-enhanced MRI (DCE MRI) peak enhancement, DCE MRI enhancement slope, DCE MRI washout slope, and apparent diffusion coefficient (ADC). Populations of progression and non-progression were compared for statistical significance reported as p-values. Categorical decision trees were generated to identify progression and assessed for receiver operating characteristic (ROC) area under the curve (AUC), sensitivity, and specificity.

Results

An example mpMRI images comparing scan 1 and scan 2 is shown in Figure 1. Superimposed on the T2W and ADC images are the PCa maps estimates of cancer lesions. The increase in the dominant lesion cancer volume at scan 2 is visible. Some scattered, small volumes exist which may be erroneous indications by the PCa maps model or other foci of cancer.
The majority of cases had increased PCa volume at scan 2 vs scan 1, as shown by most data above the 1:1 line in Figure 2. A lower bounding line to indicate erroneous outliers, as determined when the risk maps were validated³, was placed on the plot at 0.75·vol_PCa_scan1 - 1 (cc). This indicates 7 / 93 (8%) cases as outliers for the volume estimation. Figure 3 has Bland-Altman plots, showing the difference between scan 2 and scan 1 PCa volumes (Figure 3a) and fractional volumes (Figure 3b) of each case. The volumes and fractional volumes were significantly larger overall (mean increases 0.92±2.1cc and 1.3%±4.1%), and volumes increased more for larger cancers, p<0.024. At scan 2, the PCa volumes did not differ between cases with progression and those without (mean = 3.6 cc, p-value = 0.91). However, the PCa fractional volumes, normalized by the prostate volumes, did exhibit a trend to be larger in cases that progressed (9.1% vs. 7.2%, p=0.060) (Figure 4). In other words, the cancer in scan 2 tended to comprise more of the prostate in cases that progressed.
A decision tree model to identify progression was generated with sensitivity 0.84, specificity 0.57, and ROC AUC 0.75. The factors in this decision tree were PCa volume (fraction), mean DCE-MRI enhancement slope, PCa risk map mean probability within the cancer volume, and PCa volume (fraction) difference between scans.

Discussion

An increase of scan 2 volume over scan 1 was consistent with an expectation that PCa may increase in volume over time and indicates reasonable performance of the PCa volume calculation, even in a low-risk prostate cancer population on AS. Cases demonstrating decreases in cancer volume may be due to errors in the automatic volume calculation, which most commonly occur with image artifacts³. At follow-up (scan 2), cancer volumes tended to comprise a larger fraction of the prostate in cases with progression on biopsy to GG2.
Because of small differences in populations, several factors were required in decision tree models to identify progression. Sources of variability in this dataset include biopsy sampling errors and PCa map errors, which may be due to image artifacts, alignment errors, or biological differences. Even with these limitations, the good classification results indicate the potential of MRI cancer risk maps to assist in identifying progression and to aid monitoring men on AS.

Conclusion

The MRI PCa risk maps method has automated, quantitative features to provide metrics from mpMRI such as estimates of cancer volume and aggressiveness. It encompasses the entire prostate overcoming the limited sampling of biopsy. This study shows the potential for MRI PCa risk maps to enhance existing capabilities to identify prostate cancer progression during active surveillance.

Acknowledgements

Funding from California Tobacco-Related Disease Research Grant: 28IR-0060.