Ensemble of Small Classifiers For Imbalanced White Blood Cell Classification

Abstract

Automating white blood cell classification for diagnosis of leukaemia is a promising alternative to time-consuming and resource-intensive examination of cells by expert pathologists. However, designing robust algorithms for classification of rare cell types remains challenging due to variations in staining, scanning and inter-patient heterogeneity. We propose a lightweight ensemble approach for classification of cells during Haematopoiesis, with a focus on the biology of Granulopoiesis, Monocytopoiesis and Lymphopoiesis. Through dataset expansion to alleviate some class imbalance, we demonstrate that a simple ensemble of lightweight pretrained SwinV2-Tiny, DinoBloom-Small and ConvNeXT-V2-Tiny models achieves excellent performance on this challenging dataset. We train 3 instantiations of each architecture in a stratified 3-fold cross-validation framework; for an input image, we forward-pass through all 9 models and aggregate through logit averaging. We further reason on the weaknesses of our model in confusing similar-looking myelocytes in granulopoiesis and lymphocytes in lymphopoiesis. Code: https://gitlab.com/siddharthsrivastava/wbc-bench-2026.

Figures (4)

• Figure 1: Example phase 1 (unperturbed) images for each class from the WBCBench competition dataset.
• Figure 2: Number of samples per class in our expanded dataset.
• Figure 3: Model architecture. Left: our model consists of a $3$-ensemble of $3$ different architectures: a ConvNeXT-V2 Tiny 10205236, a SwinV2-Tiny Liu_2022_CVPR, and a DinoBloom-Small Koc_DinoBloom_MICCAI2024 model. In each, we replace the pretrained classification head with a simple MLP $h$. Each instantiation of an architecture is fined-tuned independently on $2$ out of $3$ folds of data and validated on the last fold to mitigate overfitting, giving 3 models for each architecture and a total of $9$ models. The final model $\mathcal{M}$ passes an input image through all $9$ architectures and averages the logits to produce the $13$-class logit vector. Right: At inference, we generate multiple augmented views of each image via random flips and rotations. These views are independently processed by $\mathcal{M}$ and their logits are averaged to obtain the final class prediction.
• Figure 4: Confusion matrices of our ensemble trained on the expanded dataset. Left: confusion matrix of $1149$ high-confidence misclassified examples, $841$ of which come from WBCBench and the rest from Acevedo20; found using confident learning northcutt2021confidentlearning. The dotted red box groups the classes in granulopoiesis and the green box captures lymphopoiesis; MO occurs in monocytopoiesis. Right confusion matrix generated using out-of-fold examples for each group of models.