White blood cell classification
April 2026
Classifying microscope images of white blood cells into 13 very imbalanced classes, from a baseline on handcrafted features to a fine-tuned ConvNeXt-Tiny with XGBoost on its embeddings, validated by 5-fold stratified cross-validation.
The problem
28,901 training images, from 13,015 segmented neutrophils down to 11 prolymphocytes. The metric is the macro-F1, so every class weighs the same and the rare ones decide the score. Several of them are successive maturation stages of the same cell line and look very much alike.
Baseline on handcrafted features
The cell is segmented by Otsu thresholding on the saturation channel, and 35 features describe its shape, color and texture. The best of four classifiers, an MLP, reaches a macro-F1 of 0.429. Its accuracy is 0.80, because the large classes are well recognized while most rare classes are almost never predicted.
Iterating on a single split
| Step | Change | Validation macro-F1 | Kaggle |
|---|---|---|---|
| V0 | ResNet-18, no rebalancing | 0.680 | |
| V6 | Class-balanced sampler, no cell crop | 0.680 | |
| V7 | Weight decay 5e-4 | 0.688 | 0.653 |
| V9 | Squeeze-and-Excitation attention | 0.650 | 0.674 |
| V10 | Cosine classification head | 0.696 | 0.629 |
| V11 | ConvNeXt-Tiny backbone | 0.676 |
The validation split and the leaderboard often disagreed: the cosine head had the best validation score and the worst submission. With some classes having only a few validation images, one split was too noisy to choose between models, which led to cross-validation.
Final pipeline and results
On each of 5 stratified folds: fine-tune ConvNeXt-Tiny with a class-balanced sampler and early stopping, extract its 768-dimensional embeddings, and train XGBoost on them. The test prediction averages the 5 folds; this final submission scored 0.696 on the official Kaggle leaderboard, the best of all submissions.
| Mean macro-F1 | Std | Min | Max | |
|---|---|---|---|---|
| ConvNeXt-Tiny | 0.7257 | 0.0138 | 0.7052 | 0.7435 |
| XGBoost on its embeddings | 0.6935 | 0.0375 | 0.6619 | 0.7566 |
The fine-tuned network alone is better and more stable. One likely reason: XGBoost learns on embeddings of the images the network was just fine-tuned on, which are better separated than those of unseen images. The submitted ensemble used XGBoost, so the better option was not the one submitted.
Limitations
The validation fold also drives early stopping, so these scores are slightly optimistic. The remaining errors are on rare and similar classes: in the best fold, 48 of the 78 band neutrophils are predicted as segmented neutrophils, their mature form.