Attention mechanisms.

The incidence of acute myeloid leukemia (AML) is increasing annually, and timely diagnostic and treatments can substantially improve patient survival rates. AML typing traditionally relies on manual microscopy for classifying and counting myeloid cells, which is time-consuming, laborious, and subjective. Therefore, developing a reliable automated model for myeloid cell classification is imperative. This study evaluated the performance of five widely-used classification models on the largest publicly available bone marrow cell dataset (BM). However, the accuracy of the classification model is significantly affected by the imbalance in the distribution of bone marrow cell types. To address this issue, this study analyzed five different Loss functions and seven different attention mechanisms. When the classification models is chosen, Swin Transformer V2 was found to perform the best. However, the lightweight model RegNetX-3.2gf had significantly fewer parameters and a significantly faster inference speed than Swin Transformer V2, and its F1 Score was only 0.032 lower than that of Swin Transformer V2. Accordingly, RegNetX-3.2gf is strongly recommended for practical applications. During the evaluation of Loss function and attention mechanism, the Cost-Sensitive Loss Function (CS) and the channel attention mechanism Squeeze-and-Excitation Networks (SE) demonstrated superior performance. The optimal model (RegNetX-3.2gf + CS + SE) achieved an average precision of 68.183%, an average recall of 63.722%, and an average F1 Score of 65.155%. This model exhibited significantly improved performance compared to the original dataset results, achieving an enhancement of 17.183% in precision and 10.655% in the F1 Score. Finally, the class activation maps demonstrate that our model focused on the cells themselves, especially on the nucleus when making classifications. It proved that our model was reliable. This study provided an important reference for the study of bone marrow cell classification and a practical application of the model, promoting the development of the intelligent classification of AML.

急性髓系白血病（acute myeloid leukemia, AML）的发病率逐年上升，及时的诊断与治疗可显著提升患者生存率。传统AML分型依赖人工显微镜对髓系细胞进行分类与计数，该过程耗时耗力且主观性较强，因此开发可靠的髓系细胞自动分类模型已迫在眉睫。本研究针对目前规模最大的公开骨髓细胞数据集（bone marrow cell dataset, BM），评估了五种主流分类模型的性能。然而，骨髓细胞类型分布不均衡的问题会显著影响分类模型的准确率。为解决这一问题，本研究分析了五种不同的损失函数（Loss function）与七种不同的注意力机制（attention mechanism）。经对比测试，Swin Transformer V2的分类性能最优，但轻量级模型RegNetX-3.2gf的参数量与推理速度均显著优于Swin Transformer V2，且其F1分数（F1 Score）仅比Swin Transformer V2低0.032。因此，实际应用中优先推荐采用RegNetX-3.2gf。在对损失函数与注意力机制的评估中，代价敏感损失函数（Cost-Sensitive Loss Function, CS）与通道注意力机制压缩与激励网络（Squeeze-and-Excitation Networks, SE）表现更优。最优组合模型（RegNetX-3.2gf + CS + SE）的平均精确率达68.183%，平均召回率为63.722%，平均F1分数（F1 Score）为65.155%。相较于原始数据集的基准结果，该模型性能实现显著提升，精确率与F1分数分别提升了17.183%与10.655%。最后，类激活图（class activation maps）结果显示，本模型在分类时聚焦于细胞本身，尤其关注细胞核区域，这证实了模型的可靠性。本研究为骨髓细胞分类研究与模型的实际应用提供了重要参考，推动了AML智能分型的发展。