Prostate Gleason SSL ViT

A Gleason classifier trained on one hospital's slides learns that hospital's staining protocol and scanner alongside the tissue; move it to another centre, another magnification, or a needle biopsy instead of a tissue microarray core, and the agreement it reported on its own test set does not survive. That fragility is the standing obstacle in automated prostate grading, where the two predominant Gleason patterns set the ISUP grade group that drives treatment.

The fix here is applied to the training distribution rather than the architecture. Five prostate cohorts — tissue microarray cores and whole-slide images, prostatectomy and needle biopsy specimens, four scanner magnifications — are pooled, Macenko colour-normalised to a shared target, and used to fine-tune one vision transformer twice: as a benign-versus-malignant classifier across 5×, 10×, 20× and 40× patches, and as a GP3/GP4/GP5 classifier over malignant tissue only. A slide is graded by running that cascade over its patches and applying the published ISUP rule to the resulting pattern percentages — arithmetic over patch calls, not a learned slide-level head, so every slide-level number traces back to countable tiles.

It was built at the Menzies Institute for Medical Research at the University of Tasmania with the pathology department at Royal Hobart Hospital and Pandani Solutions, and published in Prostate Cancer and Prostatic Diseases in March 2025. The self-supervision is inherited: the backbone's DINOv2 pretraining is Meta's, on 142 million natural images, and the adaptation to H&E is supervised. That separates it from encoders pretrained on histology itself, such as HistoEncoder and UNI, and from the group's later Pan-Cancer TP53 MIL ViT, which freezes its DINOv2 encoder beneath a learned aggregator where this one unfreezes every layer.

#Key Features

• Multi-resolution malignancy screen: The first-pass classifier trains on 5×, 10×, 20× and 40× patches together, so the benign-versus-malignant call does not assume the magnification a slide was digitised at.
• Register tokens against attention artefacts: Four register tokens absorb the high-norm activations that otherwise surface in uninformative background regions of ViT attention maps.
• Rule-based ISUP aggregation: Any pattern above 5% of malignant patches counts as substantial, and primary and secondary patterns follow established ISUP criteria rather than a second learned model, keeping the slide-level output auditable.
• Attention maps reviewed by a pathologist: Per-head attention localises the structures the grade is defined by — well-formed glands for GP3, cribriform for GP4, absent glandular architecture for GP5.

#Technical Details

The backbone is a DINOv2 vision transformer with patch size 14 and four registers: 86.6 million parameters, 115 million activations and 518×518 pixel inputs, chosen over EfficientNet and other ViT variants on SICAP. Slides are tiled into non-overlapping 512×512 patches at roughly 10× to match SICAPv2, filtered at a 5% or 20% tissue-content threshold, Macenko-normalised with staintools, and re-input at 518×518 under Fastai augmentations. Both classifiers train identically: ten epochs at a base learning rate of 2e-3 with early stopping, then all layers unfrozen for ten more under a one-cycle policy from 1e-7 to 1e-4. The corpus spans 12,711 images from 3,648 patients across eight cohorts; the classifiers were fit on five of them — SICAPv2, TMAZ, AGGC, GC2019 and a multi-resolution Singapore set — on one A100.

Internal validation gives the binary classifier AUROC 0.999, accuracy 0.985 and quadratic weighted κ 0.967 over 4,640 benign and 8,113 malignant patches, and held-out κ of 0.876–0.995; the multi-class classifier reaches κ 0.841 overall, AUROC 0.970/0.962/0.994 for GP3/GP4/GP5, and held-out κ of 0.774–0.888. Against a board-certified pathologist's consensus over 400 PANDA tiles it scores κ 0.752 across four classes, where the pathologist's own intra-observer κ was 0.802. Those held-out figures are splits of the training cohorts; on the three external cohorts — 11,132 images from 2,176 patients — slide-level agreement is materially lower: κ 0.593 for combined Gleason patterns and 0.587 for ISUP grade group on PANDA, and 0.618 and 0.620 on NADT-Prostate.

#Applications

The intended use is assisting the reporting pathologist: pre-screening specimens for malignancy, highlighting the regions carrying each Gleason pattern so a reader can confirm or overrule them, and producing a reproducible grade group where inter-observer variability is the known weak point. It reads only the H&E slide routine workup already produces.

#Impact

The contribution is generalisation engineering rather than a new architecture, and the useful result is the gap it publishes: κ near 0.59 on PANDA against κ above 0.87 on held-out splits of the training cohorts measures how much of a computational pathology headline number is cohort-specific. Other limits are on the record too: patch labels come from region-level mask annotations rather than per-patch review, GP5 is scarce everywhere, and the artifacts are closed-set classifiers over four labels for one organ. Nothing has been released — no code, no checkpoint, no hosted inference, no stated model licence — and three of the four authors declare a founding interest in a company commercialising automated histopathology assessment.