基于卷积神经网络的乳腺癌风险预测研究

文章利用卷积神经网络研究钼靶检查图像,进而对乳腺癌进行良性恶性识别。将DDSM数据库中的540张良性乳腺钼靶钙化图像、554张恶性图像进行随机旋转平移加倍扩充和高斯去噪处理。采用CNN_32/64、Resnet18_32/64和Resnet50_32/64的6个模型对乳腺钙化图像进行良性、恶性分类。结果显示CNN_64模型效果最优,测试集准确率为99.67%,验证集准确率为62.33%,AUC为59.29%。模型性能较为良好,具有一定的临床实践指导价值。

In this paper, the convolution neural network is used to study the molybdenum target image to identify the benign and malignant breast cancer. 540 benign breast molybdenum target calcification images and 554 malignant images in DDSM database were doubled and expanded by random rotation and translation, and Gaussian denoising was performed.CNN_32/64, Resnet18_32/64 and Resnet50_32/64,6 models were used to classify benign and malignant breast calcification images. The results show that the effect of CNN_64 model is the best, the accuracy of test set is 99.67%, the accuracy of verification set is 62.33%, and AUC is 59.29%. The performance of the model is good and has a certain guiding value in clinical practice.