Organizations: University of Oxford, Oxford, United Kingdom · University College London, London, United Kingdom · University of Bristol, Bristol, United Kingdom · Chulalongkorn University, Thailand
We present WBCBench 2026, an ISBI challenge and benchmark for automated WBC classification designed to stress-test algorithms under three key difficulties: (i) severe class imbalance across 13 morphologically fine-grained WBC classes, (ii) strict patient-level separation between training, validation and test sets, and (iii) synthetic scanner- and setting-induced domain shift via controlled noise, blur and illumination perturbations. All images are single-site microscopic blood smear acquisitions with standardised staining and expert hematopathologist annotations. This paper reviews the challenge and summarises the proposed solutions and final outcomes. The benchmark is organised into two phases. Phase 1 provides a pristine training set. Phase 2 introduces degraded images with split-specific severity distributions for train, validation and test, emulating a realistic shift between development and deployment conditions. We specify a standardised submission schema, open-source evaluator, and macro-averaged F1 score as the primary ranking metric.
Figures & tables
Figure 1: Geographical distribution of participants: (left) by country and (right) by continent.
Figure 2: WBCBench 2026 representative images for all 13 WBC classes (columns) at four degradation severity levels (rows: Pristine, Mild, Moderate, Extreme). Coloured borders indicate severity. PLY moderate/extreme cells are absent due to rare-class protection (no moderate/extreme degradation applied to classes with <0.5% coverage).
Split
Phase 1 train
Phase 2 train
Phase 2 eval
Phase 2 test
Patients Proportion
74 15%
222 45%
49 10%
148 30%
Num Images
8 288
24 897
5 350
16 477
Pristine
1.00
0.07
0.40
0.35
Mild
0.00
0.55
0.45
0.45
Moderate
0.00
0.30
0.12
0.15
Extreme
0.00
0.08
0.03
0.05
Table 1: Dataset partitioning and degradation severity fractions per split.
Operator
Parameter
Mild
Moderate
Extreme
Geometric, blur & noise
Crop
max ratio r
0.02–0.10
0.10–0.20
0.20–0.30
Gaussian blur
σ [px]
0.2–1.0
1.0–2.0
2.0–3.0
Motion blur
kernel k [px]
3–15
5–25
20–35
apply prob. pm
0.1–0.3
0.3–0.7
0.7–1.0
Gaussian noise
std. σn
2–8
8–18
18–25.5
Table 2: Severity-dependent parameters. Parameters are sampled uniformly; motion blur angle ∼U(0∘,360∘) ; ΔS , ΔI : additive offsets in HSV and pixel [0–255] scale, respectively.
Rank
Team
Macro-F1
Bal. Acc
Macro Prec
Macro Spec
1
FDVTS_WBC [ 15 ]
0.777
0.753
0.818
0.996
2
PathMedAI [ 12 ]
0.771
0.733
0.834
0.994
3
jht010312
0.740
0.742
0.756
0.995
4
CPRL
0.720
0.701
0.759
0.995
5
PACV
0.719
0.707
0.746
0.995
6
AIO-MHIL [ 8 ]
0.708
0.693
0.738
0.995
Table 3: Final leaderboard: top-10 results on phase2_test with organiser baselines.
Class
Support
Avg F1
Min F1
Max F1
SNE
8188
0.984
0.981
0.987
LY
4269
0.960
0.953
0.964
MO
1508
0.923
0.902
0.936
EO
492
0.968
0.949
0.979
BA
303
0.969
0.947
0.985
BL
902
0.871
0.843
0.892
Table 4: Per-class F1 averaged over the top-10 submissions, sorted by difficulty (easiest first). PLY and PC results are statistically unreliable due to very low test support.
University of Technology, Ho Chi Minh City, Vietnam · Vietnam National University, Ho Chi Minh City, Vietnam · University of Science, Ho Chi Minh City, Vietnam +2
Department of Computer Science and Engineering, The Hong Kong University of Science and Technology, Hong Kong, China · Department of Chemical and Biological Engineering, The Hong Kong University of Science and Technology, Hong Kong, China · Division of Life Science, Hong Kong University of Science and Technology, Hong Kong, China +2