Robustness of a Restriction Spectrum Imaging (RSI) quantitative MRI biomarker for prostate cancer: assessing for systematic bias due to age, race, ethnicity, prostate volume, medication use, or imaging acquisition parameters

Introduction Prostate multiparametric magnetic resonance imaging (mpMRI) has greatly improved the detection of clinically significant prostate cancer (csPCa). However, the limited number of expert sub-specialist radiologists capable of interpreting conventional prostate mpMRI is a bottleneck for universal access to this healthcare advance. A reliable and reproducible quantitative imaging biomarker could facilitate implementation of accurate prostate MRI at clinical sites with limited experience, thus ensuring more equitable patient care. Restriction Spectrum Imaging restriction score (RSIrs) is an MRI biomarker that has shown the ability to enhance the qualitative and quantitative interpretation of prostate MRI. However, patient-level factors (age, race, ethnicity, prostate volume, and 5-alpha-reductase inhibitor (5-ARI) use) and acquisition-level factors (scanner manufacturer/model and protocol parameters) can affect prostate mpMRI, and their impact on quantitative RSIrs is unknown. Methods RSI data from patients with known or suspected csPCa were collected from seven centers. We estimated effects of patient and acquisition factors on prostate voxels overall (Method 1: benign patients only) and on only the maximum RSIrs within each prostate (RSIrs max ; Method 2: benign and csPCa patients) using linear models. We then tested whether adjusting for any estimated systematic biases would improve performance of RSIrs for patient-level detection of csPCa, as measured by area under the ROC curve (AUC). Results Using both Method 1 and Method 2, we observed statistically significant effects on RSIrs of age and acquisition group (p < 0.05). Prostate volume had significant effects using only Method 2. All of these effects were small, and adjusting for them did not improve csPCa detection performance (p ≥ 0.05). AUC of RSIrs max for patient-level csPCa detection was 0.77 (95% CI: 0.75, 0.79) unadjusted, compared to 0.77 (0.76, 0.79) and 0.74 (0.72, 0.76) after adjustment using Method 1 and 2 respectively. Conclusion Age, prostate volume, and imaging acquisition factors may lead to systematic differences in RSIrs, but these effects are small and have minimal impact on performance of RSIrs for detection of csPCa. RSIrs can be used as a reliable biomarker across a wide range of patients, centers, scanners, and acquisition factors.