Characterization of prostate MRI focal lesions by combining human reading, artificial intelligence and prostate-specific antigen density: A multi-reader study.

PURPOSE: The purpose of this study was to compare, across readers with varying experience, the characterisation of prostate MRI lesions as grade group (GG) ≥ 2 cancer, by using the PI-RADS version 2.1 (PI-RADSv2.1) score alone and by combining prostate specific antigen density (PSAd), the PI-RADSv2.1 score and the output of a radiomics-based algorithm (Q-CAD).

MATERIALS AND

METHODS

The MULTI database in which 21 readers (seven experienced seniors, seven less-experienced seniors, seven juniors) had assigned a PI-RADSv2.1 score to 240 prostate MRI lesions was retrospectively used. The lesions were outlined by two independent experts to compute their Q-CAD score. For each reader, four biopsy strategies were simulated. PI-RADS3 and PI-RADS4 strategies triggered biopsy in PI-RADSv2.1 ≥ 3 and PI-RADSv2.1 ≥ 4 lesions respectively. Combined3 and Combined4 strategies triggered biopsy when at least two of the following conditions were fulfilled: positive PI-RADSv2.1 score (≥ 3 for Combined3; ≥ 4 for Combined4), positive Q-CAD score (≥ 0.45 in peripheral zone; ≥ 0.79 in transition zone), PSAd ≥ 0.15 ng/mL/cm3.

RESULTS

A total of 232 lesions were included. Using lesions' delineations by Expert 1 for the three readers' experience groups, the Combined3 strategy was significantly less sensitive for GG ≥ 2 cancers (87-88% vs. 91-96%; P = 0.026 to < 0.001), but significantly more specific (45%-55% vs. 15%-34%; P < 0.001) than the PI-RADS3 strategy. The Combined4 strategy was less sensitive than the PI-RADS4 strategy (84%-86% vs. 85%-91%) but the difference was significant only for less-experienced seniors (P = 0.023); it was significantly more specific (51%-63% vs. 27%-50%; P < 0.001) in all groups. The Combined4 strategy provided the highest net benefit for risk thresholds >12%-16%. Using lesions' delineations by Expert 2 yielded similar results.

CONCLUSION

The combined strategies significantly increased specificity, at the cost of slightly reducing sensitivity for GG ≥ 2 cancers.