基于多尺度多特征卷积神经网络的肝硬化识别

Cirrhosis recognition based on multi-scale multi-feature convolutional neural network

针对B超图像肝硬化识别中有标记数据不足和单一尺度特征缺乏鲁棒性的问题,提出了一种基于数据增强的多尺度多特征卷积神经网络M-CNN模型。首先在有限的肝硬化数据集上进行数据增强,一定程度上避免了过拟合现象。为了增强特征鲁棒性,将三种不同尺度的样本作为模型输入,模型同时学习到了不同尺度的特征,将网络不同层的多尺度信息做加权求和,使得不同层的特征对最终结果有不同影响,从而提高了模型的泛化能力。同时改进分类器权重系数,灵活调整权重灵敏度,最终在决策时提升了整个分类器的性能。实验表明,对测试集的分类准确率达到了99.2%。

In allusion to the problem of insufficient marker data and the lack of robustness of single scale features in liver cirrhosis recognition from B-mode ultrasound images,a multi-scale and multi-feature convolution neural network M-CNN model based on data enhancement is proposed.Firstly,data enhancement was performed on limit-ed cirrhosis datasets,which partially avoids overfitting.In order to enhance the robustness of features,three scales of samples were input into the model.The model learns the characteristics of different scales at the same time.The multi-scale information of different layers was weighted sum,which can make the characteristics of different layers have different effects on the final results,thus improving the generalization ability of the model.At the same time,the weight coefficient of the classifier was improved,and the sensitivity of the weight was adjusted flexibly.Finally,the performance of the whole classifier was improved in decision making.Experiments show that the classification ac-curacy of the test set reaches 99.2%.