Transfer Learning for Normal Peripheral Blood Cells Classification in Microscopy Images

Automated classification of peripheral blood cells from microscopy images offers a key tool for hematologic diagnostics, reducing the need for manual analysis while improving accuracy and efficiency. Integrating these systems into clinical workflows enables faster abnormality detection, lower diagnostic variability, and access to expert-level analysis in underserved regions. In this work, we apply transfer learning with Convolutional Neural Network (CNN) backbones- ResNet18, ResNet50, and ConvNeXt-Tiny-to classify eight types of normal peripheral blood cells from a dataset of 17,092 images. Using a stratified train-test split and 5 -fold cross-validation, we achieve test accuracy up to 98.30% and a micro-average AUC of 0.999. Misclassifications primarily occur at boundaries involving immature granulocytes, indicating potential areas for targeted improvement. Unlike prior work on this dataset, which achieved accuracies of$96-97 \%$with custom CNNs, our method uses pre-trained modern architectures to deliver better performance and efficiency. In addition, Gradient-weighted Class Activation Mapping (Grad-CAM) is used as a qualitative visualization method to compare prediction behavior across different fine-tuned pretrained models (ResNet-18, ResNet-50, and ConvNeXt).