基于改进的YOLOv3算法的乳腺超声肿瘤识别

PDF下载HTML

Title:

Tumor recognition in breast ultrasound images based on an improved YOLOv3 algorithm

作者:

徐立芳¹, 傅智杰², 莫宏伟²

1. 哈尔滨工程大学 工程训练中心，黑龙江 哈尔滨 150001;
2. 哈尔滨工程大学 自动化学院，黑龙江 哈尔滨 150001

Author(s):

XU Lifang¹, FU Zhijie², MO Hongwei²

1. Engineering Training Center, Harbin Engineering University, Harbin 150001, China;
2. Automation College, Harbin Engineering University, Harbin 150001, China

关键词:

乳腺癌; 超声影像; YOLOv3; SE-Res2Net; 下采样模块; 残差连接; 密集连接

Keywords:

breast cancer; ultrasonography; YOLOv3; SE-Res2Net; downsample block; residual connection; dense connection

分类号:

TP181

DOI:

10.11992/tis.202010004

摘要:

为了提高乳腺癌诊断的效率以及准确性，本文提出一种基于改进的YOLOv3算法来构建一个乳腺超声肿瘤识别算法，辅助医生进行乳腺癌的诊断。首先在Res2Net网络上融入SE模块构建SE-Res2Net网络来取代原始YOLOv3中的特征提取网络，以此提升模型特征提取的能力。然后通过搭建一个新型下采样模块（downsample block）来解决原始模型中下采样操作容易出现信息丢失的不足。最后为了进一步提升模型特征提取的能力，结合残差连接网络以及密集连接网络的优点构建Res-DenseNet网络来替换原始模型的残差连接方式。实验结果表明：改进后的YOLOv3算法比原始YOLOv3算法的mAP提高了4.56%，取得较好的检测结果。

Abstract:

To improve the efficiency and accuracy of breast cancer diagnoses, a breast ultrasound tumor recognition algorithm based on an improved YOLOV3 algorithm is proposed to assist doctors in breast cancer diagnosis. First, the SE module is integrated into Res2Net to construct Se-Res2Net to replace the original feature extraction network in YOLOv3 to improve the ability of model feature extraction. Then, a new Downsample Block is built to solve the problem of information loss in the downsampling operation of the original model. Finally, to further improve the ability of feature extraction, the residual connection network and dense connection network are combined to construct Res-DenseNet to replace the residual connection mode of the original model. The experimental results show that the above improvements are effective, and the mAP of the improved algorithm is 4.56% higher than that of the original algorithm.