目标检测算法在乳腺癌病灶的影像学诊断上的应用

Application of Object Detection Algorithm in Imaging Diagnosis of Breast Cancer

针对目前基于卷积神经网络的医学影像目标检测算法存在的速度不足以及精确度不够的问题,提出将SSD 目标检测算法用于乳腺癌病灶检测的实验方法。首先对现有的一万余张超声图像数据使用图像翻转、反转、高斯模糊等变换方法进行数据增强,增加样本多样性供后续网络模型进行训练,提高网络的准确率,增强网络的泛化性,提升分类器性能;下一步将样本数据输入到SSD 网络,均匀地在图片不同位置采用不同的抽样比进行密集抽样,最后利用卷积神经网络提取特征后进行分类回归。性能采用比较不同网络的mAP 来进行度量,SSD 算法的mAP 为85.09%,明显高于其他同类算法。显然基于SSD 的乳腺癌病灶检测具有较高的准确性和有效性,且具有良好的鲁棒性。

To solve the problem of insufficient speed and insufficient accuracy of medical image object detection algorithm based on convolutional neural network, proposes an experimental method for SSD object detection algorithm for breast cancer lesion detection. The method firstly uses the image flipping, inversion, Gaussian blur and other transformation methods to enhance the data of the existing 10,000 ultrasound image data, and increases the sample diversity for subsequent network model training, which improves the accuracy and enhancement of the network. The generalization of the network improves the performance of the classifier;the next step is to input the sample data into the SSD network, and uniformly sample the samples with different sampling ratios in different positions of the image. Finally, the convolutional neural network is used to extract the features and then perform classification and regression. The performance is measured by comparing the mAP of different networks. The mAP of the SSD algorithm is 85.09%, which is significantly higher than other similar algorithms. SSDbased breast cancer lesion detection has high accuracy and effectiveness, and has good robustness.