Advanced Diffusion Magnetic reSonance Prostate Imaging for Reducing Extraneous Biopsies (ASPIRE) - Calibration Component (Multi-centre Scanner Calibration) and Validation Component (Multi-centre Validation)

Calibration Component Analysis: The goal is to characterise differences in quantitative MRI metrics between scanners and derive calibration adjustments. For each pair of scanners (e.g., Philips vs Siemens), a scatter plot of VERDICT fIC values (or other

Calibration Component Analysis: The goal is to characterise differences in quantitative MRI metrics between scanners and derive calibration adjustments. For each pair of scanners (e.g., Philips vs Siemens), a scatter plot of VERDICT fIC values (or other metric) will be examined across all subjects who were scanned on both. We will calculate the mean difference and 95% limits of agreement (Bland-Altman analysis) to quantify bias and variability. A linear regression will be performed, and the regression equation (slope, intercept) will be used as the calibration formula. If needed, higher-order terms or non-linear fit will be considered, but simplicity is preferred. We will also assess the repeatability of measurements on the same scanner (some participants may have had repeat scans on the same machine or two Philips sites, etc.). The outcome of Calibration Component will be reported descriptively. No formal hypothesis test is applicable, but we will note if the inter-scanner difference

Direct comparison between VERDICT, fast VERDICT, and RSI

For each patient, we will have three outcomes (for each advanced method combined with MRI). We will compare, for example, MRI+VERDICT vs MRI+RSI specificity by looking at the subset of patients who have discordant results (one test positive, the other negative, and no cancer or cancer accordingly) using McNemar's test. We will do this pairwise for all combinations (VERDICT vs RSI, VERDICT vs fast VERDICT, RSI vs fast VERDICT). Additionally, we may compare AUC (area under the ROC curve) for the continuous metrics from each technique. The ROC curves will be generated for VERDICT fIC, RSI index, etc., against the outcome, possibly stratified by site or vendor. An averaged (pooled data) ROC comparison will indicate overall performance differences

Comparison of calibrated vs uncalibrated threshold performance

This will involve looking at the sensitivity and specificity of, for example, a uniform fIC threshold of 0.41 across all sites vs using each site's calibrated equivalent threshold (e.g. 0.35 on Siemens, etc.). We will report whether calibration improved the alignment of results. A paired comparison of these two approaches in the same dataset will be conducted (each patient can be classified by both methods). We hypothesize calibration will modestly improve specificity or consistency across sites, but this needs confirmation.

Subgroup analyses

We will examine performance by site and by scanner vendor. For each site individually, we will replicate the primary analysis to see if VERDICT added benefit at that site (these are the "individually powered" validations - each site with \~150 patients should be powered \~80% to see a similar effect size). This will tell us if any site deviates significantly. We will also aggregate data by vendor (e.g., combine both Siemens sites' data) to see if, say, all Siemens scanners behave similarly vs others. If needed, statistical models like logistic regression with site or vendor as random effects could be used to formally test for heterogeneity, but with a minimum of four sites it may be simpler to descriptively report these results.

Negative MRI cohort

Although patients with a negative MRI likely won't have biopsy verification, we will track any who had follow-up MRIs or biopsies. We will calculate the observed outcome at 1 year: how many were found to have cancer later. This is not comparative, but it gives an estimate of false negatives for both standard and advanced MRI (since an advanced MRI might conceivably pick up something in an MRI-negative case; we will see if any such case occurred and what their follow-up showed).