WBCBench2026 Classification Challenge

Single-cell blood smear image classification across 13 white blood cell types. Achieved 3rd place on the leaderboard with ~0.70 macro F1 using an ensemble approach.

Challenge Overview

The dataset presents a severe class imbalance problem:

Approach Evolution

Failed Attempt: DinoBloom Foundation Model

Initially experimented with DinoBloom, a foundation model pre-trained on pathology blood smear and bone marrow slides. Despite extensive fine-tuning, classification performance plateaued at 0.33 macro F1. UMAP visualization of CLS tokens revealed the fundamental issue: features were largely inseparable.

Successful Solution: Ensemble with Curriculum Learning

Switched to a conventional deep learning approach with three key innovations:

1. Extensive Augmentations

• H&E staining variations
• Color jitter
• Random affine transformations
• Geometric augmentations

2. Curriculum Learning for Noise Robustness

Validation set analysis revealed edge cases with intensive noise (Gaussian, salt-and-pepper). Implemented curriculum learning to gradually introduce noise and blur during training, preventing the models from being overwhelmed early on.

3. Three-Model Ensemble

• Swin Transformer
• ConvNeXt v2
• ResNet-50

All models initialized with pre-trained weights and trained using class-weighted sampling.

4. Test-Time Augmentation

Applied geometric augmentations during inference. Probability aggregation across models and augmentations proved most effective.

Results

• Final Score: ~0.70 macro F1 on test set
• Leaderboard Rank: 3rd place
• Key Finding: Curriculum learning for noise adaptation was critical for handling challenging edge cases

Alternative Approach: Synthetic Data Generation (CytoDiff)

Given the severe class imbalance—with the rarest class having only 13 images—learning generalizable features proved challenging. We explored a parallel generative approach using CytoDiff, a Stable Diffusion-based model fine-tuned with LoRA on the challenge dataset.

Key Results from Synthetic Data Augmentation (from the original paper):

• ResNet accuracy: 27% → 78% (+51%) with 5,000 synthetic images per class
• CLIP-based accuracy: 62% → 77% (+15%)

While synthetic data generation showed promise for addressing extreme imbalance, the ensemble approach with curriculum learning ultimately delivered superior performance on the challenge test set.

📁 Full implementation and analysis: CytoDiff/

Repository Structure

├── ConvNeXt_v2.py          # ConvNeXt v2 implementation
├── SwinTransformer.py      # Swin Transformer implementation
├── ResNet50.py             # ResNet-50 implementation
├── ensemble_predict.py     # Ensemble prediction script
├── figures/                # Visualizations and analysis
└── requirements.txt        # Dependencies

Dependencies

pip install -r requirements.txt