RESUMO
Differential classification of prostate cancer grade group (GG) 2 and 3 tumors remains challenging, likely because of the subjective quantification of the percentage of Gleason pattern 4 (%GP4). Artificial intelligence assessment of %GP4 may improve its accuracy and reproducibility and provide information for prognosis prediction. To investigate this potential, a convolutional neural network (CNN) model was trained to objectively identify and quantify Gleason pattern (GP) 3 and 4 areas, estimate %GP4, and assess whether CNN-predicted %GP4 is associated with biochemical recurrence (BCR) risk in intermediate-risk GG 2 and 3 tumors. The study was conducted in a radical prostatectomy cohort (1999-2012) of African American men from the Henry Ford Health System (Detroit, Michigan). A CNN model that could discriminate 4 tissue types (stroma, benign glands, GP3 glands, and GP4 glands) was developed using histopathologic images containing GG 1 (n = 45) and 4 (n = 20) tumor foci. The CNN model was applied to GG 2 (n = 153) and 3 (n = 62) tumors for %GP4 estimation, and Cox proportional hazard modeling was used to assess the association of %GP4 and BCR, accounting for other clinicopathologic features including GG. The CNN model achieved an overall accuracy of 86% in distinguishing the 4 tissue types. Furthermore, CNN-predicted %GP4 was significantly higher in GG 3 than in GG 2 tumors (P = 7.2 × 10-11). %GP4 was associated with an increased risk of BCR (adjusted hazard ratio, 1.09 per 10% increase in %GP4; P = .010) in GG 2 and 3 tumors. Within GG 2 tumors specifically, %GP4 was more strongly associated with BCR (adjusted hazard ratio, 1.12; P = .006). Our findings demonstrate the feasibility of CNN-predicted %GP4 estimation, which is associated with BCR risk. This objective approach could be added to the standard pathologic assessment for patients with GG 2 and 3 tumors and act as a surrogate for specialist genitourinary pathologist evaluation when such consultation is not available.
