Research On Liver Tumor Segmentation In CT Images Based On Fully Convolutional Networks

Liver cancer is one of the most common cancers in the world and is a serious threat to human health.The segmentation of liver tumors plays an important role in the diagnosis and treatment of liver cancer.Computed Tomography(CT)is the most frequent image modality for liver lesion assessment in clinical practice.It remains to be a challenging task to get an accurate liver tumor segmentation from abdominal CT scans automatically.Traditional image segmentation algorithms cannot achieve a good tumor segmentation performance due to the high variability in tumor shape,size,location and number of occurrences,as well as fuzzy boundaries,noise and other artifacts in CT images.In this paper,we present an automatic liver tumor segmentation method based on fully convolutional neural networks.Our contributions are three-fold: Firstly,we proposed a CT image preprocessing module that is jointly trained with the segmentation network,to find optimal preprocess settings via backpropagation,thus obviating the need of a separate and empirical CT image preprocessing procedure.Secondly,we developed a novel 2D fully convolutional neural network under the encoder-decoder framework.Our encoder is on the basis of the latest DenseNet classification network for a better feature representation ability and a higher training efficiency of deep neural networks.In the decoder part,considering the existence of multi-scale tumors,we introduced dense connections between upsample blocks in the decoder.This allows the upsample block to use semantic feature maps from all previous upsample blocks at different scales.In addition,we added U-Net like long skip connections between the encoder and decoder to combine high level semantic features and low level detail information.Through the intra-block dense connections in the encoder module,the inter-block dense connections in the decoder module and the long skip connections between the encoder and decoder,we built our main segmentation network.Finally,since the model trained with the cross entropy loss has coarse segmentation results on the tumor boundaries,we introduced the Lovász-Softmax loss function which directly optimizes the segmentation metric.We combined the two types of loss functions with dynamic weights,which means that the weights of two types of losses are linearly determined by the iteration of the training process.The hybrid loss puts more weights on the cross entropy loss in the early stage of the training process and shifts gradually to the Lovász-Softmax loss in the late stage.This leads to a coarse-to-fine liver tumor segmentation process.We achieved a fully end-to-end liver tumor segmentation framework through the design of the above three aspects.Our method was evaluated on two public liver tumor segmentation datasets.It achieved an enhanced segmentation performance on the multi-scale tumors and the tumor boundaries.To the best of our knowledge,our approach outperformed all other 2D-based methods in public on the tumor segmentation performance using the online rank metric of the LiTS challenge.