%0 Journal Article
%T 基于机器学习算法的乳腺超声图像分类研究<br>Classification Study of Breast Ultrasound Images Based on Machine Learning Algorithm
%A 汪俊豪
%A 闫士举
%A 魏高峰
%A 董天峰
%A 张涛
%A 邓爽
%J Software Engineering and Applications
%P 30-44
%@ 2325-2278
%D 2023
%I Hans Publishing
%R 10.12677/SEA.2023.121004
%X 乳腺癌已成为世界上最常见的癌症，对适龄女性进行乳腺筛查尤为重要。超声图像筛查在乳腺癌的筛查中占有主要地位。计算机辅助诊断技术可以帮助放射科医生实现快速和有效的乳腺超声图像诊断。因此，我们提出了一种全流程乳腺超声诊断系统。主要方法如下：首先基于U-Net网络对超声乳腺图像进行分割，并根据分割结果将图像划分为不同的病理区域，然后提取不同病理区域的特征并按照一定的规律得到12组特征。这12个特征组在4个不同的分类器上进行了测试，最终选出了最佳特征组和最佳分类器。结果表明，特征组NO. 12在LightGBM分类器上的性能最好。在测试集中，最终的准确度、敏感性和特异性分别为0.97、0.987和0.938，AUC达到0.962，这对于放射科医师提高诊断效率和准确性具有重要意义。<br />
Breast cancer has become the most common cancer in the world. It is particularly important to provide breast screening for women of appropriate age, and the use of ultrasound during general screening is dominant, supplemented by mammography. Computer-aided diagnostic techniques can assist radiologists in achieving rapid and efficient diagnosis of breast ultrasound images. Therefore, we propose a full-flow breast ultrasound diagnostic system. Firstly, the ultrasound breast images are segmented based on the improved U-Net network framework to obtain breast tumor regions, and the images are divided into different pathological regions according to the segmentation results, and then the features of different pathological regions are extracted and the features are combined in different permutations to obtain 12 groups of features. These 12 feature groups were tested on four different classifiers, and the best feature group and the best classifier LightGBM (Light Gradient Boosting Machine) were finally selected. The results show that feature group NO. 12 has the best performance on the LightGBM classifier. In the testing set, the final accuracy, sensitivity and specificity were 0.97, 0.987 and 0.938, respectively. The AUC reached 0.962. This is of significance for radiologists to improve the efficiency and accuracy of diagnosis.
%K 乳腺肿瘤，超声图像，U-Net，形状特征，纹理特征，LightGBM<br>Breast Tumor
%K Ultrasound Image
%K U-Net
%K Shape Features
%K Texture Features
%K LightGBM
%U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=61426