A Deep Neural Network for Analysis of Multi-class Classification of Microscopic Blood Cell Images

The classification of blood cells under a microscope plays a pivotal role in diagnosing blood-related diseases. In this research, we classify eight different types of blood cells from a dataset consisting of 17,092 RGB images using a Vision Transformer (ViT) model. To enhance the model's ability to generalize, preprocessing steps such as scaling, normalization, and data augmentation are applied to the dataset. By leveraging transfer learning, the ViT model is fine-tuned with the AdamW optimizer and Sparse Categorical Cross-Entropy loss function. The experimental results demonstrate an impressive accuracy of 98.95%, surpassing traditional CNN architectures. The confusion matrix analysis reveals minimal misclassifications, mainly occurring between cell types with similar morphologies. Although ViT achieves exceptional performance, it necessitates larger datasets for full training capacity. Future work will focus on expanding the training set to include blood cell disease types, aiming to refine the model's clinical utility. Once validated with real-world clinical data, this system could assist hematologists by providing an accurate, automated blood cell classification tool.