Hunt Alexander, Schulze Holger, Samuel Kay, Fisher Robert B, Bachmann Till T
Centre for Inflammation Research, Institute for Regeneration and Repair, The University of Edinburgh, Edinburgh, United Kingdom.
Tissues, Cells and Advanced Therapeutics, Scottish National Blood Transfusion Service, NHS National Services Scotland, Jack Copland Centre, Currie, United Kingdom.
Front Bioinform. 2025 Aug 14;5:1628724. doi: 10.3389/fbinf.2025.1628724. eCollection 2025.
The identification and classification of blood cells are essential for diagnosing and managing various haematological conditions. Haematology analysers typically perform full blood counts but often require follow-up tests such as blood smears. Traditional methods like stained blood smears are laborious and subjective. This study explores the application of artificial neural networks for rapid, automated, and objective classification of major blood cell types from unstained brightfield images. The YOLO v4 object detection architecture was trained on datasets comprising erythrocytes, echinocytes, lymphocytes, monocytes, neutrophils, and platelets imaged using a microfluidic flow system. Binary classification between erythrocytes and echinocytes achieved a network F1 score of 86%. Expanding to four classes (erythrocytes, echinocytes, leukocytes, platelets) yielded a network F1 score of 85%, with some misclassified leukocytes. Further separating leukocytes into lymphocytes, monocytes, and neutrophils, while also increasing the dataset and tweaking model parameters resulted in a network F1 score of 84.1%. Most importantly, the neural network's performance was comparable to that of flow cytometry and haematology analysers when tested on donor samples. These findings demonstrate the potential of artificial intelligence for high-throughput morphological analysis of unstained blood cells, enabling rapid screening and diagnosis. Integrating this approach with microfluidics could streamline conventional techniques and provide a fast automated full blood count with morphological assessment without the requirement for sample handling. Further refinements by training on abnormal cells could facilitate early disease detection and treatment monitoring.
血细胞的识别和分类对于诊断和管理各种血液系统疾病至关重要。血液分析仪通常进行全血细胞计数，但往往需要后续检测，如血涂片检查。像染色血涂片这样的传统方法既费力又主观。本研究探索了人工神经网络在从未染色明场图像中对主要血细胞类型进行快速、自动化和客观分类方面的应用。YOLO v4目标检测架构在包含使用微流控流动系统成像的红细胞、棘红细胞、淋巴细胞、单核细胞、中性粒细胞和血小板的数据集上进行了训练。红细胞和棘红细胞之间的二元分类实现了86%的网络F1分数。扩展到四类（红细胞、棘红细胞、白细胞、血小板）时，网络F1分数为85%，有一些白细胞分类错误。进一步将白细胞分为淋巴细胞、单核细胞和中性粒细胞，同时增加数据集并调整模型参数，得到的网络F1分数为84.1%。最重要的是，在供体样本上进行测试时，神经网络的性能与流式细胞术和血液分析仪相当。这些发现证明了人工智能在未染色血细胞高通量形态分析方面的潜力，能够实现快速筛查和诊断。将这种方法与微流控技术相结合可以简化传统技术，并提供一种无需样本处理即可进行形态学评估的快速自动化全血细胞计数。通过对异常细胞进行训练进一步优化，可以促进疾病的早期检测和治疗监测。
