Fig. 4: Morphological spatial heterogeneity with deep learning-based Gleason grading.

a, Example output from the automated Gleason classifier, with primary and secondary pattern assessment. b,c, Examples of the Gleason Morisita assessment. Cells identified as epithelial cells by the cell classifier are subdivided into Gleason grades using the region’s automated Gleason segmentation. Regions with high segregation of patterns (b) will be assigned a low Gleason Morisita index, whereas regions with high mixing between Gleason grades (c) will be assigned a high Gleason Morisita index. d, Participants with greater within-section heterogeneity of Gleason pattern, as assessed by Gleason Morisita index, are associated with a shorter time to recurrence (two-sided log-rank test, χ² = 8.33, d.f. = 1, P = 0.0039; imaging cohort, n = 250 participants). e, CPH model of time to recurrence using clinical covariates and the Gleason Morisita index (imaging cohort, n = 250 participants, *P < 0.05, **P < 0.01, ***P < 0.001). The forest plot shows 95% CI of HRs and the covariate P values, derived from a Wald test. HR for the Gleason Morisita index represents the increase in hazard between the 5th and 95th percentile values (within the imaging cohort). f, ISUP grade group as a predictor of time to recurrence. A comparison is shown for the grade groups assessed by the original reporting pathologist, the reviewing pathologist for the trial and the automated classifier (imaging cohort, n = 250 participants). Grade groups are calculated from the assessed primary and secondary patterns, according to the 2014 ISUP criteria. Only the automated Gleason assessment was able to stratify the participants by time to recurrence (two-sided log-rank test, χ² = 9.52, d.f. = 3, P = 0.023). g, Confusion matrices showing the pairwise agreement of the ISUP grade groups reported by the original reporting pathologist, the reviewing pathologist for the trial and the automated classifier (imaging cohort, n = 250 participants).

Source data