Smartphone‑based portable microscopy for high‑accuracy on‑device white blood cell classification

Accurate classification of white blood cell (WBC) type and subtypes is critical for diagnosing infections, leukemia, and other conditions. However, manual counting is slow and error-prone, particularly in teaching laboratories and when done by students with little experience. We introduce an accessory-free, on-device smartphone system and method that classifies six WBC categories, including the five main types, as well as band form neutrophil and segmented neutrophil distinction. The method can be used on virtually any standard microscope and iOS device with no additional tools. A dataset of 900 stained peripheral-blood images (150 per class) was collected. A 3-level Otsu thresholding algorithm was used to isolate the nuclear band, decrease the variance in colors of microscopic images, and decrease the variance in staining. The middle level was used as input for a ResNet-18 backbone model. The model achieved a 97% overall accuracy, 100% accuracy for the five main WBC types, and 91.7% accuracy for the band-form and segmented neutrophil discrimination. The model was converted with CoreML and embedded in an iOS application (“WBC Smart Lab”). The converted model occupies 22.38MB and has an inference latency of 33 ms on an iPhone 13 Pro Max. A pilot study on smartphone-captured images shows promising real-world application. Overall, this work proposes an accessible, accurate, and practical “second opinion” for medical technology education and clinical use for WBC classification.