EfficientNet-B3-Based Multi-Class Blood Cell Classification for Automated Hematological Diagnosis

Abstract

Accurate classification of blood cell types is essential for diagnosing hematological disorders such as leukemia. Traditional manual methods are time-consuming and prone to inter-observer variability, prompting the need for automated, reliable solutions. This study presents a deep learning framework using a fine-tuned EfficientNet-B3 model for automated classification of six blood cell types from microscopic images. Using a publicly available Kaggle dataset of 12,500 labeled images, the model was trained and evaluated using precision, recall, F1-score, and accuracy. Fine-tuned EfficientNet-B3 achieved a classification accuracy of 99%, outperforming benchmark models including RetinaNet, VGG-16, a custom CNN, and BloodCell-Net. The architecture demonstrated rapid convergence, robust generalization, and minimal misclassification, particularly between morphologically similar cell types. The results underscore the model’s potential for real-time deployment in clinical settings, offering a computationally efficient and highly accurate diagnostic aid. Future work will focus on external validation, interpretability through explainable AI, and deployment on low-resource platforms.