Automated Bone Marrow Cell Classification for Haematological Disease using Machine Learning

Accurate classification of bone marrow cells is essential for diagnosing hematological disorders. However, manual classification is labor-intensive and susceptible to human error. Previous approaches, including CNN combined with SVM, CNN with XGBoost, and Siamese networks, faced challenges related to limited accuracy and poor generalization. This study utilizes transfer learning to automate and enhance the classification process. A dataset containing 170,000 expertly annotated bone marrow cell images was used to evaluate the performance of four pretrained deep learning models—VGG16, ResNet50, InceptionV3, and EfficientNetB5. These models were enhanced with additional layers and trained using data augmentation techniques. Among them, EfficientNetB5 achieved the highest performance, attaining a training accuracy of 98.47% and a validation accuracy of 89.39%, outperforming the other models. The results indicate that transfer learning models can significantly improve diagnostic accuracy and efficiency. This study provides a robust solution for real-time, automated hematological analysis, addressing limitations in current diagnostic practices.