Deep Learning (DL) Algorithms to Predict Prostate Cancer Progression on Low/intermediate-Risk Needle Core Biopsies.

335 Background: Gleason Grade Group (GG) remains the standard of care to guide disease management in prostate cancer. Deep Learning (DL) algorithms assessing GG on whole slide images (WSI) have been largely validated by comparison to pathologist-determined GG. However, few studies have assessed DL grading of needle biopsies for predicting relevant clinical end-points. Methods: WSI of biopsies from two previously published cohorts (PMID 30279579, 27693448) were analyzed by a modified DL system for Gleason grading (PMID 35233002). In the Active Surveillance (AS) cohort (case-control design), all patients (n=137) had an initial biopsy with GG1 disease. Cases (63/137) had grade reclassification to GG2+ on repeat biopsies, or GG3+ or non-organ-confined disease by radical prostatectomy (RP), while the controls (74/137) continued on AS for >8 years without grade reclassification. For the Grade Group 2 (GG2) cohort, all patients (n=285) had GG2 disease on biopsy followed by RP, and progression was defined as biochemical recurrence (38/285). One representative H&E slide containing up to two biopsy cores from the diagnostic or confirmatory biopsy in both cohorts was graded by the DL system, and evaluated by 2-3 blinded uropathologists to provide consensus Gleason grading. Results: The agreement of GG between uropathologists and the modified DL system (κ quad ) was 0.32 for the AS cohort and 0.37 for the GG2 cohort. In the AS cohort, upgrading of the initial GG1 biopsy to GG2+ by the uropathologists occurred in 10% of cases compared to 7% of controls (p = 0.78). Upgrading by the DL system to GG2+ occurred in 33% of cases compared to 8% of controls (p=0.0005). In the GG2 cohort, biochemical recurrent after RP was not significantly associated with upgrading of the original biopsy to GG3+ by the uropathologists (HR= 0.55; 95% CI: 0.21-14.88; p=0.58) but significantly associated with upgrading by the DL system (HR=4.47; 95% CI: 1.32-15.09; p=0.01) in multivariable models including pre-operative PSA, clinical stage and fraction cores involved by tumor. Conclusions: In low or intermediate risk needle biopsies (GG1 and GG2), the agreement of grading between uropathologists and modified DL system is lower than that seen in analyses on a broader range of grade groups (PMID 35233002). Nevertheless, the DL system grading outperformed uropathologist consensus in predicting progression in patients with low/intermediate grade prostate cancer receiving active surveillance or radical prostatectomy.