Machine Learning Based Normal White Blood Cell Multi-Classification Optimization

Abstract

Identification of various types of white blood cells (WBCs) in a timely accurate manner is essential for diagnosing infections, allergic diseases, and blood disorders via complete blood count (CBC) workflow. Conventional identification relies on manual microscopic analysis of Leishman- or Giemsa-stained blood smears. While accurate when done by an expert, this cumbersome, subjective process relies heavily on trained hematologists who are often scarce in many high-volume or low-resource labs. This study reports the design, implementation, and comparative testing of a deep-learning-based tool used to automatically identify five types of WBCs in blood smears: Basophil, Eosinophil, Lymphocyte, Monocyte, and Neutrophil. Three convolutional neural networks were trained and evaluated in similar conditions; two transfer-learning models, ResNet50V2 and DenseNet121 pre-trained on ImageNet with multiple phases of training and a Custom CNN. All models were trained using a two-stage approach, starting with head-only training with frozen backbone and finishing with selected fine-tuning of identified backbone top layers and data augmentation, batch normalization, dropout, early stopping, and learning-rate scheduling. Interactive desktop graphical user interface was developed to enable one-image classification, compare models’ side-by-side, and visualize dataset statistics and model performance in an exploratory manner. The results generated from the experiments performed on the held-out test set indicate that ResNet50V2 achieved the best performance in overall terms scoring 94.88% accuracy in its test and macros. This is higher than DenseNet121 performance of 88.89% and Custom CNN of 85.34%. Basophil was found to be the most unambiguous one to classify, while Eosinophil and Monocyte were ranked as the most difficult classes indicating that particular grades of classification difficulty are bound with meanings that mark certain groups of granulocytes.