Automatic segmentation of tumors combining lung prior and synergetic deep supervision
There are two challenges in automatic segmentation of complex lung tumors(CLT)on computed tomography(CT)images:1)The class indistinction between tumors and adjacent tissues;2)Intra-class inconsistencies within tumors.In order to solve these two challenges,the semantic context prior of the relationship between lung tumor and lung is proposed to be incorporated into the segmentation model,so that the model can learn the semantic context features,and the segmentation of CLT can be reconsidered from a macro perspective.The anatomical prior of lung shape is modeled using information entropy.The proposed novel attention module is embedded in the three-classified U-Net network,so as to guide the training process through domain-specific knowledge.In addition,a boundary enhancement auxiliary network was designed to obtain tumor boundary structure and maintain the consistency of tumor internal features.On this basis,a collaborative deep supervision network framework(CLT-ASegNet)was developed,which further improved the discriminant ability and convergence speed of the model by using hybrid multi-scale semantic feature fusion.CLT-ASegNet was evaluated on CLTCT1 segmentation datasets and Lung16 datasets.The experimental results show that the proposed CLT-ASegNet can effectively segment lung tumors.
attention mechanismcomplex lung tumor segmentationsemantic context priorsynergetic deep supervision
NETL
NSTL
万方数据
