High performance white blood cell classification using ensemble CNNs with Monte Carlo dropout and conformal prediction for clinical uncertainty quantification

Automated classification of white blood cell (WBC) images may support cell-morphology workflows, but high accuracy alone does not establish clinical readiness. We developed an Uncertainty-Guided Deep Ensemble (UGDE) combining MobileNetV3-Large, EfficientNet-B0, and ShuffleNetV2 with Monte Carlo dropout, temperature scaling, split conformal prediction, and Grad-CAM. The primary dataset contained 16,027 single-cell images from 78 patients (33 with leukemia and 45 without leukemia). All stated development partitions were grouped by patient, and augmentation was restricted to the proper training subset. On the held-out internal test set (3,205 images from 16 patients), UGDE achieved approximately 99% image-level accuracy and macro-F1 of 0.99. Expected calibration error was 2.3% before and 0.8% after temperature scaling. Internal empirical conformal coverage was 94.2%, 89.7%, and 79.7% at nominal coverage levels of 95%, 90%, and 80%, respectively; these slight shortfalls are reported without claiming that the targets were met. External evaluation used 2000 Raabin-WBC images after mapping the nine model outputs to five shared leukocyte categories and yielded 94% descriptive accuracy. Grad-CAM examples were used as post-hoc visualizations and were not interpreted as causal explanations. These retrospective results support further evaluation of uncertainty-aware WBC image classification, but they do not demonstrate performance on whole slides, automated differential counts, prospective patients, or clinical workflows.