Advanced Multi-Class Bone Marrow Classification: EfficientNetB5 Techniques for Better Outcomes
Vishnu Kant, Kanwarpartap Singh Gil, Sonal Jain Malhotra, Swati Devliyal · 2024
Classifying bone marrow images into multiple categories presents challenges such as class imbalance, high intra-class variability, and the need for highly precise and interpretable diagnostic models. These issues complicate the development of reliable automated clinical diagnostic systems. Accurate classification is crucial for diagnosing various hematological disorders, yet achieving this requires overcoming obstacles like ensuring result precision and interpretability, dealing with significant morphological differences, and addressing class imbalance. This study explores the application of the advanced convolutional neural network, EfficientNetB5, to address these challenges in bone marrow image classification. By employing customized loss functions, advanced data augmentation, and techniques for model interpretability, our approach achieves an impressive accuracy of $\mathbf{9 6 \%}$. These results demonstrate EfficientNetB5’s potential to enhance diagnostic accuracy and reliability in clinical settings.