基于卷积神经网络和多信息融合的三维乳腺超声分类方法

乳腺肿瘤超声图像的自动分类对于提高医生的工作效率和降低漏诊率具有十分重要的意义。新型的三维乳腺超声数据包含更多的可用于诊断的信息,但由于超声成像机理导致不同方向上的图像表现不同。针对该种乳腺超声数据,利用卷积神经网络结构的灵活性和自动学习的特性,提出3种改进的卷积神经网络模型,使其分别可以接受横截面图像输入、横截面和冠状面的双图像输入、图像和文本信息同时输入,并研究不同信息的融合对于提升乳腺肿瘤自动分类准确率的影响。在研究中,采用880幅图像(良性401幅,恶性479幅)及其标注信息进行5折交叉验证实验,得到各模型的准确率及AUC。实验结果表明,设计的模型可以适应图片与文本信息的输入,多信息融合的模型比只接受图像输入的模型准确率提升2.91%,达到75.11%的准确率和0.829 4的AUC。这些模型的提出,为多信息融合的卷积神经网络分类应用提供参考。

Classification of Breast Mass in 3D Ultrasound Images with Annotations Based on Convolutional Neural Networks

The automatic classification of breast tumor in ultrasound images is of great significance to improve doctors' efficiency and reduce the rate of misdiagnosis. The novel 3D breast ultrasound data contains more information for diagnosis, but images from different directions have their distinct performance as a result of this ultrasound imaging mechanism. For this breast ultrasound data, this paper designed three kinds of convolutional neural network model using its flexibility and characteristic of learning automatically, and the three models were able to accept transverse plane images, transverse plane and coronal plane images, images and annotations information. The effects of different information fusion on the accuracy of breast tumor classification were investigated. A dataset contains 880 images (i.e., 401 benign images, 479 malign images) and their annotations were employed, and we performed 5-fold cross validation to calculate the accuracy and AUC of each model. The experimental results indicated that the models designed in this paper can deal with the images and annotations simultaneously. Compared with the single-input model, the multi-information fusion model improved the accuracy of classification by 2.91%, and achieved the accuracy of 75.11% and AUC of 0.829 4. The proposed models provided a reference for the classification application of convolutional neural networks with multi-information fusion.