基于全卷积神经网络的肝脏CT影像分割研究

Fully convolutional neural network for liver segmentation in CT image

针对腹部CT影像邻近器官对比度较低及因个体肝脏形状差异较大等引起肝脏分割困难的问题,提出了全卷积神经网络肝脏分割模型。首先通过卷积神经网络提取图像深层、抽象的特征,再通过反卷积运算对提取到的特征映射进行插值重构后得到分割结果。由于单纯进行反卷积得到的分割结果往往比较粗糙,因此,在反卷积之前,先融合高层与低层的特征,并且通过增加反卷积的层数、减少反卷积步长,得到了更为精确的分割结果。与传统卷积神经网络的分割方法相比,该模型可以充分利用CT影像的空间信息。实验数据表明该模型能够使腹部CT影像肝脏分割具有较高的精度。

Abdominal CT images cover problems such as low contrast in adjacent organs and various performance inshape.A liver segmentation model based on fully convolutional neural network is proposed.Firstly,the deep and abstractfeatures of the image are extracted by convolutional neural network.Then interpolated reconstruction is performedthrough deconvolution operation on the extracted feature map to obtain segmentation results.Due to the simple deconvolutionacquiring segmentation results are usually rough.Before deconvolution,it applies characteristics mergence to upperand lower layers,increases the deconvolution-layer amount and reduces deconvolution-step size on the model,then getsaccurate segmentation results.Compared to convolution neural network,this model can fully use the spatial informationof CT images.Experimental results demonstrate,this model can segment abdominal liver region in CT images and reachmuch higher accuracy.