Abstract
METHODS
Two hundred and ninety-one operated prostate cancer men with pre-treatment MRI and whole-mount prostate histology were retrospectively included. T2-weighted, apparent diffusion coefficient (ADC) and fractional blood volume maps from 1.5/3T MRI systems were used. 592 histological GP3, GP4Crib- and GP4Crib+ regions were segmented on whole-mount specimens and manually co-registered to MRI sequences/maps. Radiomics features were extracted, and an erosion process was applied to minimize the impact of delineation uncertainties. A logistic regression model was developed to differentiate GP4Crib+ from GP3/GP4Crib- in the 465 remaining regions. The differences in balanced accuracy between the model and baseline (where all regions are labeled as GP3/GP4Crib-) and 95% confidence intervals (CI) for all metrics were assessed using bootstrapping.
CONCLUSION
The radiomics MRI-based model, trained on Gleason sub-patterns segmented on whole-mount specimen, was able to differentiate GP4Crib+ from GP3/GP4Crib- patterns with moderate accuracy. The most dominant feature was the 90th percentile ADC. This exploratory study highlights 90th percentile ADC as a potential biomarker for cribriform growth differentiation, providing insights into future MRI-based risk assessment strategies.
RESULTS
The logistic regression model, using the 90th percentile ADC feature with a negative coefficient, showed a balanced accuracy of 0.65 (95% CI: 0.48-0.79), receiver operating characteristic area under the curve (AUC) of 0.75 (95% CI: 0.54-0.92), a precision-recall AUC of 0.35 (95% CI: 0.14-0.68).
OBJECTIVE
To differentiate cribriform (GP4Crib+) from non-cribriform growth and Gleason 3 patterns (GP4Crib-/GP3) using MRI.