Gleason Grading

The system defines five primary histological grades (1–5) based on glandular architecture, not nuclear atypia. Pattern 1 shows well-formed, uniform glands. Pattern 2 has more stromal separation. Pattern 3 displays infiltrative, variably sized glands. Pattern 4 exhibits fused, cribriform, or poorly formed glands. Pattern 5 shows solid sheets, cords, or single cells with comedonecrosis. Deep learning models must learn to distinguish these subtle architectural transitions.

The final Gleason Score is the sum of the two most prevalent patterns: the primary (dominant) and secondary (sub-dominant) patterns. Scores range from 6 (3+3) to 10 (5+5). A tertiary pattern (highest grade, <5% of tumor) is reported separately. Automated systems must perform patch-level classification and aggregate spatial statistics to compute the score, often using Multiple Instance Learning (MIL) to handle the gigapixel scale.

The 2014 ISUP Grade Group system simplifies Gleason scores into five prognostic groups to better reflect clinical risk:

• Grade Group 1: Gleason 3+3=6
• Grade Group 2: Gleason 3+4=7
• Grade Group 3: Gleason 4+3=7
• Grade Group 4: Gleason 4+4=8, 3+5=8, 5+3=8
• Grade Group 5: Gleason 9–10 This stratification is the standard output for AI diagnostic models.

Within Pattern 4, distinguishing cribriform glands (solid nests with punched-out lumina) from glomeruloid structures (tufted, glomerulus-like) is critical. Cribriform morphology is an independent predictor of biochemical recurrence and metastasis. High-performance models use attention mechanisms to focus on these diagnostically salient regions, often visualized with Grad-CAM saliency maps for pathologist verification.

Manual Gleason grading suffers from significant inter-observer variability, particularly at the 3+4 vs. 4+3 boundary. Studies show kappa values as low as 0.43 for community pathologists. Deep learning models trained on consensus annotations from expert genitourinary pathologists aim to provide a standardized, reproducible reference. Uncertainty quantification techniques flag ambiguous cases for mandatory human review.

While not part of the Gleason grade itself, perineural invasion (PNI) — cancer tracking along nerves — is a critical contextual finding. AI models performing Gleason grading are often extended to simultaneously detect PNI as a multi-task learning objective. The spatial relationship between high-grade patterns and nerve bundles provides a powerful composite prognostic signal for treatment planning.