| Bone marrow erythroid cells are among the most critical blood cells in the human body.The erythroid cells’ number,a key indicator for the health status of the human body,is essential for the prevention of anaemia and other serious blood diseases.It also provides a vital reference for patient treatment and rehabilitation.The detection and classification of erythroid cells rely heavily on the traditional manual detection method,which requires considerable human,material,and financial resources and introduces subjective differences in detection and counting.Therefore it often leads to inferior detection and classification performance.However,the deep learning technique can solve the drawbacks mentioned above and improve the performance of erythroid cells detection and classification efficiently and effectively.Most current cell detection and classification research focus on white blood cell（WBC）data.However,relatively few studies concentrate on the automated detection and fine-grained classification of erythroid cells.Besides,compared with WBCs,bone marrow erythroid cells at different maturation stages are more densely distributed and have minimal morphological differences.Therefore,it makes erythroid cell detection and fine-grained classification more challenging.This paper has collected and constructed the bone marrow erythroid cell（BMEC）dataset in collaboration with the Second Hospital of Jilin University.And then,we propose a new pipeline to accomplish the detection and fine-grained classification on the BMEC dataset.The main contents are as follows:1.This study collected and constructed a large scale bone marrow erythroid cell dataset to comprehensively analyze the characteristics of erythroid cells and improve the performance of detection and classification.This dataset contains four types of cells,239 whole field erythroid cell images and 5,666 individual cell images.2.To solve the issues of dense distribution and mutual obscuration of erythroid cells,this paper demonstrates a single-stage cell detection pipeline based on the RetinaNet model,which can accomplish efficient detection and achieve an accurate localization of four types of erythroid cells.3.This paper proposes a novel attention mechanism called shape attention based on the cell shape mask images extracted from the RGB cell images.The shape attention module,a flexiable adapter worked with the state of the art backbone networks,can improve the fine-grained classification performance for the erythroid cell.Experiments demonstrate that the single-stage erythroid cell detection scheme proposed in this paper finally achieves a detection accuracy of 90.70%,which is 0.6%better than the manual detection by blood experts.In addition,the novel shape attention module can enhance the cell shape information and improve the fine-grained classification.After applying the shape attention module,several classification metrics on BMEC dataset are improved（Accuracy:+0.37%,Precision:+1.79%,Recall:+0.70%,F1-Score:+0.78%）.Besides,the fine-grained classification approach outperforms the state of the art methods on two open-sourced WBCs datasets. |
