采用DoubleUNet网络的结直肠息肉分割算法

Colorectal polyp segmentation algorithm using DoubleUNet network

由于结直肠息肉的大小、颜色和质地各异,且息肉与周围粘膜的边界不清晰,导致息肉分割存在较大挑战。为提高结直肠息肉的分割准确率,本文提出了一种改进的DoubleUNet网络分割算法。该算法首先对息肉图像进行去反光处理,并通过数据扩增方法将训练数据集进行扩大;接着,在DoubleUNet网络的解码器部分引入注意力机制,并将网络中的空洞空间卷积池化金字塔(ASPP)模块替换为密集连接空洞空间卷积池化金字塔(DenseASPP)模块,以提高网络提取特征的能力;最后,为提高小目标的分割精度,提出利用Focal Tversky Loss函数作为本算法的损失函数。该算法在Kvasir-SEG、CVC-ClinicDB、ETIS-Larib、ISIC和DSB数据集测试中的准确率分别为0.9530、0.9642、0.8157、0.9503和0.9641,而DoubleUNet算法在上述数据集的准确率分别为0.9394、0.9592、0.8007、0.9459和0.9496。实验结果表明本文算法相对于DoubleUNet算法具有更好的分割效果,可以有效的辅助医师切除结直肠异常组织从而降低息肉癌变的概率,且能够应用于其它医学图像分割任务中。

Colorectal polyps are different in size,color and texture,and the boundaries between the polyps and the surrounding mucosa are not clear,leading to significant challenges in polyp segmentation.In order to improve the segmentation accuracy of colorectal polyps,this paper proposes an improved DoubleUNet network segmentation algorithm.The algorithm first de-reflects the polyp image,and the training dataset is amplified by data-augmentation method;then introduces an attention mechanism in the decoder part of the DoubleUNet network,and replaces the atrous spatial pyramid pooling module of the network with a densely connected atrous spatial pyramid pooling module to improve the ability of the network to extract features;finally,in order to improve the segmentation accuracy of small targets,the Focal Tversky Loss function is proposed as the loss function of this algorithm.The accuracies of the algorithm in the KvasirSEG,CVC-ClinicDB,ETIS-Larib,ISIC,and DSB dataset are 0.9530,0.9642,0.8157,0.9503,0.9641,respectively,while the accuracies of the DoubleUNet algorithm in the above datasets are 0.9394,0.9592,0.8007,0.9459,0.9496.The experimental results show that the algorithm in this paper has a better segmentation effect than the DoubleUNet algorithm,which can effectively assist physicians to remove abnormal tissues of colorectum and thus reduce the probability of cancerous polyps and it can be applied to other medical image segmentation tasks as well.