SOSNet:一种非对称编码器-解码器结构的非小细胞肺癌CT图像分割模型

SOSNet:An Asymmetric Encoder-Decoder Structure Model forAutomatic Segmenting Non-Small Cell Lung Cancer CT Images

非小细胞肺癌严重损害人类健康,早期非小细胞肺癌CT(Computed Tomography)图像中的肿瘤结节体积小,不易发现,极易造成漏诊和误诊.为了精确分割非小细胞肺癌CT图像中的小体积肿瘤结节,本文提出SOSNet(Small Object Segmentation Networks)自动分割模型,利用ResNet(Residual Network)基础层和空洞卷积构造非对称编码器-解码器结构作为分割主网络,利用轴向取反注意力模块ARA(Axial Reverse Attention)逐步擦除背景中对分割有影响的结构,再使用结构细化模块SR(Structure Refinement)对主网络输出的粗略特征图进行结构细化,从而实现非小细胞肺癌肿瘤结节分割.在非小细胞肺癌公开数据集的实验测试表明,本文提出的小目标自动分割模型SOSNet可以有效分割出非小细胞肺癌CT图像中的小体积肿瘤结节,其mDice(mean-Dice)、mIoU(mean Intersection over Union)、Sensitivity、F1、Specificity、平均绝对误差MAE(Mean Absolute Error)均优于当前最先进的小目标分割模型CaraNet(Context Axial Reverse Attention Network).

Non-small cell lung cancer(NSCLC)will imperil human health seriously.The tumor nodules at the early stage of NSCLC are so small that it is very difficult to detect them in the CT(Computed Tomography)images,which will easily lead to the missed diagnosis and misdiagnosis of NSCLS.To automatically segment the small tumor nodules in CT images of NSCLC accurately,the SOSNet(Small Object Segmentation Networks)model is proposed.The ResNet(Residu⁃al Network)base layer and the dilated convolution are adopted to construct the asymmetric encoder-decoder structure to be the segmentation main network of SOSNet.The ARA(Axial Reverse Attention)module is adopted to gradually erase those structures which may influence the segmentation results from the background.Then the SR(Structure Refinement)module is used to refine the rough feature maps outputted by the main network,so as to achieve the segmentation for NSCLC tumor nodules.Experimental results on the open access NSCLC datasets demonstrate that the proposed SOSNet model can effec⁃tively segment small volume tumor nodules in CT images of NSCLC.It is superior to the state-of-the-art small object seg⁃mentation model of CaraNet in terms of mDice(mean Dice),mIoU(mean Intersection over Union),Sensitivity,F1,Speci⁃ficity and MAE(Mean Absolute Error),respectively.