Skip to main content
Back

Oxford's WBCBench 2026 Exposes Where Blood Cell AI Still Dangerously Fails

An ISBI 2026 challenge pushed 241 teams to classify 13 white blood cell types under class imbalance and synthetic scanner noise, with foundation models topping the leaderboard.

Pro

  • WBCBench 2026 launched at ISBI 2026: 55,012 blood smear images across 13 fine-grained WBC classes from 493 patients
  • Challenge tests robustness under class imbalance, patient-level splits, and synthetic scanner-induced domain shift
  • 241 teams registered, 101 submitted; winner reached macro-F1 0.777 vs ResNet-50 baseline at 0.635
  • Top methods converge on DinoBloom foundation models with LoRA and heavy ensembling
  • Rare cells remain unsolved: plasma cells averaged F1 0.15, band neutrophils 0.42 across top ten teams
  • Dataset now public on Hugging Face with open evaluator

Automated white blood cell classification looks solved on paper. Public benchmarks routinely report near-perfect accuracy, and vendors sell digital hematology analyzers into hospital workflows. Those numbers hide a messy reality: models trained on small, balanced datasets tend to collapse the moment they see a slightly different stain, a noisier scanner, or a rare blast cell they were never taught to recognize. WBCBench 2026, an ISBI 2026 challenge, was built to stress-test that gap between benchmark scores and clinical deployment.

The organizers, from Oxford, Bristol, UCL and Chulalongkorn University, released a single-site dataset of blood smear images and ran a competition designed to expose every place current models tend to break. From 241 registered teams and 101 valid submissions, 7 teams achieved macro-F1 above 0.70 and the top team outperformed the ResNet-50 baseline by 14.2 macro-F1 points. Even the winners struggled on the rarest cell types, arguably the more interesting result.

Why WBC classification breaks in the wild

White blood cell morphology is central to diagnosing leukemia, myelodysplastic syndromes, and severe infections. In practice, hematologists inspect Wright-Giemsa stained peripheral blood smears and count neutrophils, lymphocytes, monocytes, eosinophils, basophils, and blasts. The manual workflow is slow and suffers from inter-observer variability, but the automated alternatives have their own problems: commercial analyzers like the Sysmex DI-60 still require heavy expert review, and their accuracy drops precisely on the abnormal cells that matter most for leukemia diagnosis.

The core issue is that most published models are evaluated on cohorts with weak patient-level separation, few blast examples, and almost no variation in acquisition conditions. WBCBench 2026 closes that gap with a clinically realistic single-site dataset covering 13 expert-annotated WBC classes including blasts and rare subtypes, strict patient-level separation with group-stratified sampling to preserve minority coverage, a controlled degradation pipeline that injects scanner-like variability, and a standardised evaluation protocol based on macro-averaged F1 with an open-source evaluator and leaderboard.