Lightweight, Stain-Normalized White Blood Cell Classification with Explainability and Web-Deployed Prototype

Accurate white blood cell (WBC) classification from stained microscopy images is critical for hematological diagnosis but is impeded by stain variability and class imbalance. This work aims to deliver a lightweight, stain-robust, and explainable WBC classifier with a publicly available demo to support reproducible evaluation. We normalize input images using Vahadane stain normalization to reduce color heterogeneity and apply targeted data augmentation to address class imbalance. An EfficientNetV2-B1 backbone (≈7 million parameters) is trained on the augmented dataset, and model decisions are interpreted using Grad-CAM heatmaps. Statistical robustness is quantified through bootstrap resampling with 1,500 samples to generate 95% confidence intervals for accuracy. The resulting model attains 98.69% test accuracy, with bootstrap analysis indicating a 95% confidence interval of 98.22–99.14%, and Grad-CAM visualizations corroborate that predictions rely on biologically plausible regions. We deploy an interactive web application that displays the original image, the Vahadane-normalized image, Grad-CAM overlays, and per-class confidence scores to promote transparency and practical use.