检验医学 ›› 2025, Vol. 40 ›› Issue (10): 998-1003.DOI: 10.3969/j.issn.1673-8640.2025.10.012

• 论著 • 上一篇    下一篇

基于YOLOV8深度学习模型辅助诊断急性早幼粒细胞白血病

罗远美1, 刘畅1, 陈立龙1, 李振彰2, 岳雨彪2, 李杨2, 徐令清1()   

  1. 1 广州医科大学附属清远医院 清远市人民医院检验医学部广东 清远 511500
    2 广州医科大学生物医学工程学院广东 广州 511436
  • 收稿日期:2024-05-24 修回日期:2024-09-20 出版日期:2025-10-30 发布日期:2025-11-07
  • 通讯作者: 徐令清,E-mail:lingqing_xu@126.com
  • 作者简介:罗远美,女,1983年生,学士,副主任技师,主要从事感染和血液系统疾病人工智能应用研究。
  • 基金资助:
    广东省医学科学技术研究基金项目(A2021490);清远市科技计划基金项目(221104197683799);清远市人民医院医学科研基金项目(20190209);清远市人民医院医学科研基金项目(202301-201);清远市人民医院医学科研基金项目(202301-318)

Acute promyelocytic leukemia diagnosis assisted by YOLOV8 deep learning model

LUO Yuanmei1, LIU Chang1, CHEN Lilong1, LI Zhenzhang2, YUE Yubiao2, LI Yang2, XU Lingqing1()   

  1. 1 Department of Clinical LaboratoryQingyuan People's Hospital,the Affiliated Qingyuan Hospital of Guangzhou Medical UniversityQingyuan 511500,Guangdong, China
    2 School of Biomedical EngineeringGuangzhou Medical UniversityGuangzhou 511436,Guangdong, China
  • Received:2024-05-24 Revised:2024-09-20 Online:2025-10-30 Published:2025-11-07

摘要:

目的 分析YOLOV8深度学习(DL)模型快速识别急性早幼粒细胞白血病(APL)细胞图像的可行性。方法 收集2004—2023年广州医科大学附属清远医院56例明确诊断为APL且未经治疗的患者的骨髓涂片,用MGview1.0软件进行拍摄,采集其中1 551张清晰图像,采用Make Sense软件标注细胞类别(“ZYLXB”类别和“Other”类别)。共标注异常早幼粒细胞图片5 041张和其他类型细胞图片1 606张,随机抽取4 033张异常早幼粒细胞图片和1 330张其他类型细胞图片作为训练集,训练结束后的剩余图片纳入验证集。建立YOLOV8深度学习模型,评价该模型快速识别APL细胞图像的效能。结果 YOLOV8 DL模型识别异常早幼粒细胞的准确率达91%,且随置信度阈值的增加,精确度也呈逐渐提高的趋势,在阈值约为0.9时达到高点。结论 YOLOV8 DL模型可对APL图像进行快速且准确的识别。

关键词: YOLOV8深度学习, 急性早幼粒细胞白血病, 骨髓涂片

Abstract:

Objective To analyze the feasibility of using the YOLOV8 deep learning(DL) model to rapidly identify cell images of acute promyelocytic leukemia(APL). Methods Bone marrow smears from 56 untreated patients diagnosed with APL at the Affiliated Qingyuan Hospital of Guangzhou Medical University from 2004 to 2023 were collected. Medical image capture software was used to take pictures,and 1 551 clear images were collected and labeled with cell types(ZYLXB and Other) by Make Sense software. A total of 5 041 abnormal promyelocyte images and 1 606 other-type cell images were labeled. Totally,4 033 abnormal promyelocyte images and 1 330 other-type cell images were randomly selected as the training set,and the remaining images after training were included in the validation set. A YOLOV8 DL model was established to evaluate the model's ability to rapidly identify APL cell images. Results The accuracy of the YOLOV8 DL model in identifying abnormal promyelocytes reached 91%,and the precision gradually increased with the increase of the confidence threshold,reaching a peak at approximately 0.9. Conclusions The YOLOV8 DL model can rapidly and accurately identify APL images.

Key words: YOLOV8 deep learning, Acute promyelocytic leukemia, Bone marrow smear

中图分类号: