💻 computer science

WBCBench 2026: A Challenge for Robust White Blood Cell Classification Under Class Imbalance

This paper introduces WBCBench 2026, an ISBI challenge and benchmark designed to evaluate the robustness of automated white blood cell classification algorithms against severe class imbalance, strict patient-level data separation, and realistic domain shifts caused by imaging perturbations.

Original authors: Xin Tian, Xudong Ma, Tianqi Yang, Alin Achim, Bartłomiej W Papież, Phandee Watanaboonyongcharoen, Nantheera Anantrasirichai

Published 2026-04-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xin Tian, Xudong Ma, Tianqi Yang, Alin Achim, Bartłomiej W Papież, Phandee Watanaboonyongcharoen, Nantheera Anantrasirichai

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where doctors have to look at thousands of tiny blood cells under a microscope every day to diagnose diseases like leukemia. It's a tedious, eye-straining job, and even the best experts can get tired or make mistakes. To help, scientists are building AI "digital assistants" to do the counting and sorting for them.

But here's the problem: most of these AI assistants are like students who only study for a test using perfect, textbook examples. They ace the exam when the pictures are crystal clear, but the moment the lighting is dim, the camera is shaky, or the image is blurry, they panic and fail.

WBCBench 2026 is a new, tough "final exam" designed to fix this. It's a global competition created by researchers from universities like Oxford and Bristol to see which AI can actually handle the messy, imperfect reality of the real world.

Here is how the challenge works, explained through a few simple analogies:

1. The "13-Category Sorting Game"

Imagine you are a librarian trying to sort 13 different types of very similar-looking books.

The Easy Books: You have thousands of copies of "Neutrophils" and "Lymphocytes" (the most common blood cells).
The Rare Books: You have only a handful of copies of "Blasts" or "Plasma Cells" (rare cells that often signal serious illness).
The Trap: Most AI models are trained to just sort the thousands of common books perfectly. They ignore the rare ones because they don't see them often. But in a hospital, missing those rare books could mean missing a cancer diagnosis. WBCBench forces the AI to pay attention to every book, even the rare ones.

2. The "Two-Phase Test"

The competition is split into two rounds to trick the AI into being honest:

Phase 1 (The Practice Run): Participants get a set of perfect, high-definition photos. This is like studying in a quiet library with good lighting. Everyone learns the basics here.
Phase 2 (The Real World): This is where the real test happens. The organizers take the photos and intentionally "ruin" them. They add:
- Blur: Like taking a photo while running.
- Noise: Like static on an old TV.
- Bad Lighting: Like taking a picture in a dark room.
- Color Shifts: Like looking at the cells through a tinted filter.

The goal? To see if the AI can still identify the cells correctly even when the image looks like it was taken with a shaky hand in a dim room.

3. The "Patient Privacy" Rule

In many previous tests, an AI might cheat by memorizing the specific "face" of a patient's cells. If it saw Patient A's cells in the training set, it would recognize them in the test set, even if it didn't actually learn what the cell type was.

WBCBench 2026 has a strict rule: No cheating. The AI is trained on one group of patients and tested on a completely different group of patients it has never seen before. It has to learn the shape of the cell, not the identity of the patient.

4. The Results: Who Won?

Over 240 teams from around the world signed up (from students to big tech companies). 101 teams actually finished the race.

The Baseline: The researchers set up a "standard" AI (like a high school student) to see how hard the test was. It scored about 63% accuracy.
The Winners: The top teams (like "FDVTS WBC" and "PathMedAI") used advanced techniques. They treated the AI like a team of experts:
- One expert looks at the big picture.
- Another zooms in on tiny details.
- A third specialist is hired just to find the rare, tricky cells.
- They even used "self-training," where the AI guesses on the blurry images and then learns from its own mistakes.

The best team scored 77.7%, which is a huge improvement, but the paper admits: It's still not perfect.

5. The Big Takeaway

Even the smartest AIs are still struggling with the rarest cells (like "Blasts" and "Plasma Cells") when the images are very blurry.

Why does this matter?
This competition isn't just about winning a trophy. It's a reality check. It shows us that while AI is getting better, we can't just plug it into a hospital tomorrow. We need to build systems that are robust enough to handle bad photos, rare diseases, and messy data before we can trust them with patients' lives.

In short: WBCBench 2026 is a stress test for medical AI, proving that to save lives, our digital doctors need to be as tough and adaptable as the real ones.

1. Problem Statement

The paper addresses the critical gap between high-performing deep learning models on ideal datasets and their failure in real-world clinical settings. Specifically, it targets three major challenges in automated White Blood Cell (WBC) classification:

Severe Class Imbalance: Clinical data is naturally skewed, with mature cells (neutrophils, lymphocytes) dominating and rare, clinically critical subtypes (blasts, plasma cells) appearing infrequently.
Patient-Level Data Leakage: Many existing benchmarks fail to separate data by patient, allowing models to "memorize" patient-specific features rather than learning generalizable morphology.
Domain Shift and Degradation: Real-world medical imaging suffers from variability in scanners, lighting, focus, and noise, which current models often fail to handle robustly.

Current commercial systems (e.g., Sysmex DI-60) and academic models struggle with rare classes and atypical presentations, particularly under degraded conditions.

2. Methodology: The WBCBench 2026 Benchmark

The authors propose a two-phase challenge designed to stress-test algorithms under realistic constraints.

A. Dataset Construction

Source: 55,012 microscopic images from 493 patients, acquired at a single institution using standardized Wright–Giemsa staining.
Classes: 13 morphologically fine-grained WBC classes, including segmented neutrophils (SNE), blasts (BL), plasma cells (PC), and rare subtypes like prolymphocytes (PLY).
Annotation: Labels were assigned by 5 expert hematopathologists with consensus sessions for ambiguous cases.
Splitting Strategy: Strict patient-level separation is enforced. No patient appears in both training and test sets to prevent information leakage.
- Phase 1 (Training): 15% of patients (74 patients, 8,288 images). Images are pristine (no degradation).
- Phase 2 (Validation/Test): 85% of patients. Images undergo controlled synthetic degradation.

B. Degradation Pipeline

To simulate real-world acquisition variability, Phase 2 images are subjected to four severity levels: Pristine, Mild, Moderate, and Extreme.

Operators: Includes geometric transformations, Gaussian/Motion blur, Gaussian/Poisson noise, and color/illumination shifts (saturation, gamma, brightness, vignetting).
Severity Logic: Severity is determined by both the number of operators applied (1–5) and the intensity of their parameters.
Rare Class Protection: To ensure rare classes (like PLY) are not completely lost to degradation, specific rules enforce a minimum number of pristine images for every class, even in the test set.

C. Evaluation Protocol

Primary Metric: Macro-averaged F1 Score across all 13 classes. This metric penalizes models that ignore rare classes, even if overall accuracy is high.
Tie-Breaking: Balanced Accuracy $\rightarrow$ Macro Precision $\rightarrow$ Macro Specificity.
Baselines: Two standard models were provided as reference:
1. ResNet-50: Pretrained on ImageNet, fine-tuned with class-weighted cross-entropy.
2. Swin-Tiny: Vision Transformer adapted for 224×224 inputs.

3. Key Contributions

First Robustness Benchmark for Hematology: Unlike previous datasets that focus on balanced, pristine images, WBCBench 2026 explicitly integrates class imbalance, patient-level separation, and controlled domain shift into a single standardized evaluation.
Realistic Clinical Simulation: The degradation pipeline mimics scanner and setting-induced artifacts (blur, noise, lighting) rather than random noise, providing a more rigorous test of model robustness.
Open-Source Ecosystem: The challenge provides a standardized submission schema, an open-source evaluator, and a public leaderboard to facilitate reproducible research.
Community Engagement: The challenge attracted 241 teams globally, with 101 submitting valid results, demonstrating significant community interest in robust medical AI.

4. Results

Performance Overview:
- Baselines: ResNet-50 achieved a Macro-F1 of 0.635; Swin-Tiny achieved 0.643.
- Top Performers: The top team (FDVTS WBC) achieved a Macro-F1 of 0.777, a significant improvement of ~14 points over the baseline.
- General Trend: 73% of participants outperformed the ResNet-50 baseline, and 65% beat the Swin-Tiny baseline. The median score was 0.656.
Top Techniques:
- Ensemble Methods: Top teams utilized ensembles of foundation models (e.g., DINOv3, DinoBloom, Swin-Small).
- Rare Class Handling: Successful strategies included dedicated binary classifiers for rare classes (e.g., PLY), contrastive embedding heads (MedSigLIP), and self-training via pseudo-labeling.
- Preprocessing: Techniques like GAN-based denoising (Pix2Pix), stain normalization (Macenko), and cell-specific filters (e.g., "spikiness score") were common.
- Loss Functions: Use of Focal Loss, LogitAdjustment, and class-balanced sampling to mitigate imbalance.
Per-Class Analysis:
- Easy Classes: Mature cells (SNE, LY, EO, BA) achieved F1 > 0.96.
- Hard Classes: Rare and transitional cells remained difficult.
  - Plasma Cells (PC): F1 = 0.15 (severely underrepresented, n=15).
  - Band Neutrophils (BNE): F1 = 0.42 (morphologically similar to SNE/Myelocytes).
  - Promyelocytes (PMY): F1 = 0.50.
- Key Insight: High micro-averaged accuracy does not guarantee robustness; top models succeeded by specifically boosting recall on rare classes without sacrificing performance on abundant ones.

5. Significance and Implications

Clinical Relevance: The benchmark highlights that while deep learning has advanced, robustness to degradation and imbalance remains the primary bottleneck for deploying automated morphology analysis in routine hematology, especially for leukemia diagnosis where rare blasts are critical.
Future Directions: The results suggest that future research must move beyond simple fine-tuning. Effective solutions require domain-adaptive foundation models, specialized pipelines for rare classes, and synthetic data augmentation that respects clinical constraints.
Resource Availability: By releasing the dataset, evaluator, and baselines, the authors provide a crucial resource for the community to continue developing and testing robust medical AI systems, moving the field closer to real-world deployment.

In conclusion, WBCBench 2026 establishes a new standard for evaluating medical image classification, proving that current state-of-the-art models still struggle with the "long tail" of rare diseases and real-world image quality variations.