Liver Lesion Segmentation Based On Convolutional Neural Network

Medical imaging diagnosis is a common method to detect organ diseases.In general,it is necessary for experienced doctors to make judgment by artificial observation.However,the accuracy of manual diagnosis is limited by the doctor's ability.At the same time,the large number of images also brings great pressure to the interpreting.In recent years,with the rapid development of the Neural Network technology,more and more computer vision tasks have made great progress by using the Convolutional Neural Network(CNN).In this paper,the automatic liver segmentation based on CNN was studied.The following tasks were mainly completed:Based on the UNet network structure,we designed a volume full convolutional network structure for medical image data.Considering that medical image data is a kind of fuzzy and complex three-dimensional structure for organ edge features,some special structure designs such as skip connection,dilated convolution,down-sampling and deconvolution were used in this network.Secondly,in order to avoid the inaccuracy of segmentation caused by categories imbalance of dataset,the DICE loss function was used in network optimization.Finally,in order to improve the generalization ability of the volume full convolutional network,the strategy of fine-tuning the original network with actual data set was adopted.On the basis of liver segmentation research,we further made a segmentation study of liver lesion.For liver lesion segmentation,the hybrid densely connected UNet network was proposed.Considering that the liver lesions are generally small-sized and morphologically diverse three-dimensional structure,we first designed a three-dimensional densely connected UNet network,which used UNet,dense connection,down-sampling,deconvolution and other structures.However,the train stage is long and complex in time for three-dimension structure.Then,the idea of pre-extracting features by two-dimensional densely connected UNet network was adopted,and the features extracted from the twodimensional network were used as the input of three-dimensional densely connected UNet networks.Finally,in order to make full use of the extracted two-dimensional features(intraslice features)and three-dimensional features(inter-slice features),a feature hybrid network structure combining two-dimensional features and three-dimensional features was proposed.In this structure,the special transformation mechanism of data feature channel and data batch size was used to realize the mutual transformation of two-dimensional data and threedimensional data.In order to verify the effectiveness of the algorithm in this paper,some experiments were carried out on the Li TS(Liver Tumor Segmentation Challenge)data set and actual data set.For the volume full convolutional network,the Li TS training data set was used in training stage,while the test stage adopted the Li TS test set and actual data set.For the fine-tuning the network,actual training data set was used in training stage and the testing data set was used in testing stage.The experiment results showed that the volume full convolutional network is 0.6% higher than the state-of-the-art 3D UNet in DICE segmentation metric,and the DICE segmentation metric of the fine-tuned model is 1.6% higher than the original in actual testing data set.For the hybrid densely connected UNet lesion segmentation network,the training network used the Li TS training set,and the test network used the Li TS testing set and actual testing data set.This network adopted a phased training approach in train stage.The experimental results showed that the segmentation metric of the hybrid densely connected UNet network is at least 4.8% higher than other networks.