Deep learning-based bone marrow cytology classification: A solution to class imbalance

Classification of bone marrow cells is important for diagnosing hematopoietic diseases such as leukemia and lymphoma. Traditional methods, such as complete blood count and peripheral smear analysis, focus mainly on mature blood cells. However, bone marrow analysis is critical to understanding all stages of blood cell development. Manual bone marrow analysis is time-consuming, error-prone and requires expertise. In addition, the classification of bone marrow cells is complicated due to changing cell lineages and morphological variation, leading to data imbalance. This study introduces a new Embedding-Space Re-sampling Technique (ESRT) integrated into a feature extractor model to address data imbalance and improve classification. This algorithm generates synthetic samples for minority classes in the embedding space instead of relying on image data pixel space. The ESRT approach improves computational efficiency and expands decision boundaries by focusing on challenging samples to classify near decision boundaries and extract key embeddings. This method enhances the model’s ability to distinguish between classes by highlighting adversarial examples from opposing classes. Using a large 21-class bone marrow cytology dataset, the proposed framework achieved an impressive classification accuracy and Matthews Correlation Coefficient (MCC) of 89.90% and 73.19%, respectively, surpassing existing methods. Also, the accuracy and F1-score of the inference on the unseen test dataset are 75.58% and 76.11%, respectively. This framework provides a solution to data imbalance with significant efficiency, increasing model and classification performance without the need for extensive preprocessing.