BanffNET

Peritubular capillaritis is a continuous biological process — leukocytes accumulate in more capillaries, and in greater numbers, as rejection progresses. The Banff classification asks a pathologist to compress that into one of four integers, then asks a pathologist in another hospital to land on the same integer. They frequently do not. Interobserver variability in the harder lesions is the long-standing weak point of transplant histopathology, and it propagates into every diagnosis and trial endpoint built on the scores.

BanffNET refuses the integers. It keeps the pathologist's ordinal labels as its only supervision, but predicts a continuous, calibrated probability for each lesion using an aggregation step chosen to match how that lesion is defined. Banff lesions come in two geometries: some score severity — the worst single structure anywhere on the slide, such as tubulitis or intimal arteritis — while others score extent, the fraction of cortex affected, such as interstitial inflammation or fibrosis. BanffNET models the first with Noisy-OR Instance Pooling, whose probabilistic OR over patches saturates when any one region is convincingly involved, and the second with Gated Instance Pooling, which averages patch probabilities under a learned relevance gate. The operator, not a learned attention mechanism, is where the histological reasoning lives.

It was built by a European consortium led by the Systems Pathology Research Group at Leiden University Medical Center, drawing slides from eight transplant centres in the Netherlands, Belgium, Austria, Germany and Italy, and posted as a preprint in September 2026.

#Key Features

• Lesion-matched probabilistic pooling: Noisy-OR pooling for focal severity lesions and gated pooling for diffuse lesions, so the aggregation operator encodes what the Banff score measures.
• Continuous, calibrated output: slide-level scores run 0–1 rather than snapping to Banff's ordinal bins, preserving sub-threshold injury that grading discards, and are calibrated by post-hoc logistic scaling.
• Pixel-level explanations: Generic Attention Explainability runs through the frozen backbone and the lesion head to localise the tissue responsible for each score, rather than producing the diffuse patch-level heatmaps typical of attention MIL.
• Extremely small heads: each lesion head carries 1,025 (Noisy-OR) or 2,050 (gated) parameters, over 2,500-fold fewer than the attention baselines it outperforms, and scores a biopsy in about 18 seconds on one 24 GB GPU.
• Triple-stain late fusion: separate models are trained per stain and combined, so H&E, periodic acid–Schiff and Jones' silver each contribute to the biopsy-level score.

#Technical Details

Whole-slide images are masked with an in-house UNet++/MobileNetV2 tissue segmenter, tessellated into non-overlapping 256×256-pixel patches at 0.5 µm/pixel, Reinhard-normalised per stain, and encoded into 1,024-dimensional vectors by a frozen UNI backbone — the ViT-L/16 histology foundation model pretrained with DINOv2 on over 100 million tiles. Only the aggregation heads are trained, one per lesion per stain, giving 51 models across 17 lesions — the 15 Banff-defined scores plus global and focal segmental glomerulosclerosis. Training used 7,533 WSIs from Amsterdam UMC, LUMC and UMC Utrecht; the fixed system was then applied to 12,687 WSIs from Leuven, Vienna, Aachen, Nijmegen and Parma, spanning different scanners, staining protocols and case mixes. Against ABMIL (788,226 parameters) and TransMIL (2,671,633 parameters) on the identical feature pipeline, the 1,025- and 2,050-parameter heads scored higher on the four lesions tested. In a reader study where 67 renal pathologists from 24 countries each scored the same 36 biopsies, BanffNET correlated with the 67-reader mean more strongly than the individual readers did, across all lesions and independent of years of experience.

#Applications

The immediate use is reproducible quantification of transplant biopsies: standardised lesion scores for multicentre trials, retrospective scoring of archived slide collections at a scale no reader panel could cover, and a second opinion on lesions with known low rater reliability. Because the outputs are continuous they also serve as covariates in prognostic models — Cox models for death-censored graft failure built on BanffNET scores matched or exceeded pathologist-score models on external cohorts — and as a bridge to molecular diagnostics, retrieving biopsies with similar MMDx archetypes, transcriptomes and proteomes better than pathologist scores do.

#Impact

BanffNET's more interesting claim is not that it beats pathologists but that it is validated against things pathologists cannot see. Its scores correlate more closely with MMDx molecular lesion probabilities and enrich more broadly for alloimmune transcript signatures than the human grades used to train it, which argues that continuous modelling recovers tissue injury the ordinal gold standard truncates. It also reproduces the population-level correlation structure among lesions better than individual readers, despite being trained on each lesion independently, and two follow-up preprints from the same consortium already build on it. The limits are real: supervision is anchored to imperfect pathologist labels, the cohorts are European academic centres, the feature extractor has little kidney-specific training, no code or weights are released, and the work is not yet peer reviewed.