Myelodysplastic syndrome (MDS) is a heterogeneous malignant clonal disease derived from hematopoietic stem cells, characterized by abnormal myeloid cell development and a high risk of progression to acute myeloid leukaemia. Bone marrow cell examination is the main diagnosis method, which relies heavily on the pathologist’s skill and experience. However, the process is tedious and time-consuming, and the diagnostic results are subjective. Computer-aided diagnosis still faces challenges due to high intra-class and low inter-class variances of bone marrow cell. In this paper, we propose a fine-grained bone marrow cell classification that combines an adaptive mask branch module with the LGBCE loss function. The adaptive mask branch module dynamically adjusts the generation and application of masks, enabling the model to learn features more effectively during training, thereby enhancing generalization ability. The LGBCE loss function penalizes confusing categories to further improve the model’s ability to distinguish between different classes of cells. The experimental results demonstrate that our proposed method can effectively improve the accuracy of bone marrow cell classification.