基于改进型YOLO网络目标检测算法在乳腺肿瘤超声图像检测中的应用研究  

Study on the application of YOLO algorithm based on improved YOLO network in the detection of ultrasound image for breast tumor

作　　者：杨涛 杨岚兰 杨米扬 黄棋 叶双雨 付丽媛 赵红佳[3] Yang Tao;Yang Lanlan;Yang Miyang;Huang Qi;Ye Shuangyu;Fu Liyuan;Zhao Hongjia(The First Clinical Medical College,Fujian University of Traditional Chinese Medicine,Fuzhou 350004,China;Department of Radiology and Diagnosis,Fuzhou 350025,China;Department of Ultrasound,Affiliated People's Hospital of Fujian University of Traditional Chinese Medicine,Fuzhou 350004,China)

机构地区：[1]福建中医药大学第一临床医学院,福州350004 [2]中国人民解放军联勤保障部队第九〇〇医院放射诊断科,福州350025 [3]福建中医药大学附属人民医院超声科,福州350004

出　　处：《中国医学装备》2024年第9期23-27,共5页China Medical Equipment

摘　　要：目的:实现基于改进型YOLO网络目标检测算法(YOLO算法)模型对乳腺肿瘤超声图像的检测方式优化升级。方法:选取Kaggle数据库上659幅乳腺肿瘤图像作为初始数据集,采用图像标注工具Labelimg对图中检测目标进行预标注,依照7∶3的比例将659幅图像中的629幅图像划分为训练集与验证集,其余30幅图像作为测试集,对原YOLO算法引入卷积块注意力模块(CBAM)与双向特征金字塔网络(BiFPN)进行结构化改良并命名为YOLOv5-BiFPN-CBAM。将训练集与验证集置入原YOLO算法模型与YOLOv5-BiFPN-CBAM模型进行训练并经200轮迭代训练,将所得最佳权重文件用于测试集的最终化检验。结果:两种模型经过200轮迭代训练后,经验证集检验,两种模型对所有乳腺肿瘤超声图像检测的平均精度均值分别为72.1%和80.5%,将改良模型的最佳权重文件经测试集测试,改良模型相较于原始模型对图像中小目标的检测能力得到显著提升。结论:改进型YOLO算法模型与原YOLO算法模型相比,具有更高的对图像的识别度,同时提高了对乳腺肿瘤超声图像中小目标识别的精度与灵敏度,有助于提高临床中乳腺肿瘤的诊断效能。Objective:To realize the optimization and upgradation of the detection method of you only look once(YOLO)algorithm model based on the improved YOLO network on the ultrasound image for breast tumor.Methods:A total of 659 images of breast tumor of the Kaggle database were selected as the initially dataset,and the image annotation tool Labelimg was used to conduct pre-labeling for the detection targets in the images.According to a ratio as 7:3,629 images of the 659 images were divided into the train set and validation set,and the other 30 images were used as the test set.The convolutional block attention module(CBAM)and bidirectional feature pyramid network(BiFPN)were introduced into the original YOLO algorithm to underwent structural improvement,which was named as YOLOv5-BiFPN-CBAM.Both the train set and validation set were placed in original YOLO algorithm model and YOLOv5-BiFPN-CBAM model to conduct train,which included 200 rounds of iterative training.The obtained optimal weight files were used in the final test of test set.Results:After 200 rounds of iterative train for two kinds of models,the test results of validation set indicated that the mean values of average precision of two kinds of models were respectively 72.1%and 80.5%for all ultrasound images of breast tumor.The result,that the optimal weight file of improved model was tested by test set,indicated the test ability of improved model was significantly enhanced than that of original model for small target in image.Conclusion:Compared with the original YOLO algorithm model,the improved YOLO algorithm model has higher recognition capability for image,which also enhances precision and sensitivity in identifying small targets of ultrasound images of breast tumor.This model is helpful to improve the diagnostic efficiency in clinical practice for breast tumor.

关 键 词：人工智能(AI) 目标检测 乳腺肿瘤 乳腺超声 YOLO算法 

分 类 号：R445.1[医药卫生—影像医学与核医学]