基于深度Q学习和可变形卷积U-Net的肝脏肿瘤分割方法

Liver tumor segmentation method based on deep Q learning and deformable convolution U-Net

针对传统卷积神经网络U-Net对早期肝脏肿瘤的分割精度低的问题,在U-Net的基础上提出了基于深度Q学习和可变形卷积U-Net的肝脏肿瘤分割方法。首先使用深度Q学习对图像进行肿瘤目标定位,然后对目标肿瘤区域使用可变形卷积的U-Net进行分割,最后实现了粗剪裁到细分割的两段式学习框架。实验结果表明,利用该方法在肝脏肿瘤数据集上测试,其分割结果的Dice系数能够达到68%,较传统的卷积神经网络U-Net精度提升了6.89个百分点。

Aiming at the problem of low segmentation accuracy of the traditional convolutional neural network U-Net for early liver tumors,based on U-Net,a liver tumor segmentation method based on deep Q learning and deformable convolutional U-Net was proposed. Firstly,deep Q learning was used to locate the tumor target in the image. Secondly,the target tumor region was segmented using a deformable convolution U-Net. Finally,a two-stage learning framework from rough cutting to fine segmentation was realized. Experimental results show that by using this method to test on the liver tumor dataset,the DSC(Dice Similarity Coefficient)of the segmentation result can reach 68%,which is 6. 89 percentage points higher than that of U-Net.