Comparing Gleason pattern 4 measurement approaches on prostate biopsy using machine learning: a proof-of-principle study

There is increasing evidence that length of Gleason Pattern (GP) 4 on biopsy is a stronger predictor of oncological outcomes than Grade Group (GG) or %GP4. However, there is no consensus regarding the optimal method for quantifying GP4 length. We propose that machine learning (ML) can be used to quantify GP4 on digitized biopsy slides using multiple measurement approaches, allowing direct comparison of their prognostic performance. In this study, we attempt to demonstrate proof-of-principle for this approach.

We assembled a convenience sample of 726 patients with GG2-4 prostate cancer on systematic biopsy who underwent radical prostatectomy between 2014 and 2023. ML algorithm (PAIGE-AI) quantified GP4 using different approaches, varying by how gaps between cancer foci were handled. GP4 extent was assessed using linear and pixel-based metrics. Discrimination of each quantification approach, along with GG, was assessed for adverse radical prostatectomy pathology and biochemical recurrence.

We identified 15 different quantification approaches and observed differences between their discrimination. The highest discrimination was in the pixel-counting method (AUC 0.648). GP4 quantification outperformed GG for predicting adverse pathology (AUC 0.627 versus 0.608). The amount of GP3 was non-predictive once GP4 was known. These findings were consistent for BCR.

We were able to measure slides using 15 distinct measurement approaches and replicated prior findings using ML to quantify GP4. Our findings support the use of ML as a research tool to compare different GP4 quantification approaches. We intend to use our method on larger cohorts to determine which measurement approach best predicts oncological outcome.