Loading…
Vision transformer that regresses Ki-67-positive and -negative nuclei counts in breast histopathology and scores the index from H&E slides alone.
No providers recorded yet. Browse all providers
Scoring the Ki-67 proliferation index means a pathologist counting at least 500 tumour nuclei on an immunohistochemistry slide and calling each one brown or blue. Where that threshold sits is a judgement, and it moves between observers — which matters most in the 15–25% band, exactly where the index decides whether a patient is offered chemotherapy. The stain is also an extra assay: the H&E section was already cut and scanned before anyone ordered Ki-67.
This model attacks both problems by refusing to detect cells at all. A patch goes in and two numbers come out — positive nuclei and negative nuclei — so the network fixes its own staining cut-off during training instead of inheriting one, and never has to separate touching nuclei the way a segmentation pipeline must. Reaching an unstained slide takes a second step, because the IHC-trained regressor does not transfer: applied straight to H&E patches it scores R² = −46.40 on positive counts, worse than predicting the mean. The fix is cross-stain pseudo-labelling — the IHC regressor's outputs become surrogate labels for the registered H&E patches of the same tissue, and a second checkpoint trains on H&E alone, so a model that never sees a brown nucleus learns which morphology accompanies one.
The system was built at Pandani Solutions with the Menzies Institute for Medical Research at the University of Tasmania and the pathology department at Royal Hobart Hospital, posted to medRxiv in April 2025 and published in Breast Cancer Research in November 2025. It shares a backbone and a lab with Pan-Cancer TP53 MIL ViT, and sits in the same morphology-to-molecular lane as Hist2Cell and Path2Omics while reading a single protein marker at patch resolution rather than a transcriptome.
The backbone is DINOv2 ViT-Base/14 with four register tokens — 86 million parameters, 115 million activations, 518×518 inputs — with the classification head replaced by a fully connected regression head. Training fits the head for up to 20 epochs, then unfreezes every layer for 20 more under a 1-cycle schedule with discriminative rates from 1e-6 to 1e-4, using Adam, batch size 32 and an MSE loss on a single 40 GB GPU. The IHC regressor saw 3,561 annotated patches from SHIDC-B-Ki-67 and BCData, expanded to 14,244 by multi-scale augmentation, and reached MAE 9.13 and R² 0.90 on validation, above 0.90 for both cell counts on the held-out test set and 0.93 for the patch-level index. The cross-modality model, trained on 20,656 paired patches from IHC4BC, reaches R² 0.95 for positive and 0.97 for negative counts; trained on the index alone it reaches R² 0.80, and at the 30% clinical cut-off its high/low calls agree with the IHC-derived ones on 96.8% of 4,132 test patches. The gradability classifier, trained on 21,366 patches balanced across ACROBAT, TCGA-UT and the two annotated Ki-67 sets, is error-free on validation and on a 399-patch unseen set including SICAPv2. The pipeline was run over 1,996 ACROBAT whole-slide images — 1,153 H&E and 843 IHC.
The intended use is a Ki-67 report generated from the H&E slide already on the scanner: a slide-level index, per-patch counts, and a hotspot outline in place of the pathologist's manual choice of field. That removes an antibody, a block and a turnaround from the workflow, which matters most where IHC capacity is the bottleneck. The pipeline also scores genuine IHC slides, so it can act as a second reader on the conventional assay.
The paper's most useful contribution is a negative result made productive: a stain-specific counter collapses across modalities, and pseudo-labelling across registered pairs recovers it — a recipe that extends to any marker with paired IHC and H&E material. The limits are equally on the record. The cross-modality model's supervision is another model's output rather than pathologist ground truth, and the authors state that it still needs validation against annotated H&E whole-slide images, which do not publicly exist; the external runs on ACROBAT and DeepSlides are correlational, since those slides carry no Ki-67 labels. The work is breast-only and single-marker, and nothing has been released — no code, no checkpoints, no hosted inference and no stated model licence, with four of the five authors declaring a founding interest in the company commercialising it.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.