Research On Liver CT Image Segmentation Based On Deep Learning

Liver is the largest substantial organ in the human body.Liver is rich in blood vessels,complex in structure,and has many types of high-morbidity diseases,which have seriously affected human health and life.Liver surgery is one of the most commonly used treatments for liver disease.It is necessary to segment from CT images for processing and analysis to obtain pathological,physical and anatomical information,so as to provide a theoretical basis for the formulation of surgical plans.Usually,the contour of liver is manually drawn by experienced experts based on prior knowledge,which is time-consuming,labor-intensive and subjective.Therefore,researching fully automatic liver segmentation is the primary task of liver surgery treatment,which has important practical significance and application value.In recent years,many researchers have proposed semi-automatic or fully automatic liver CT image segmentation methods,which have achieved good results.However,there are still many problems remaining.For example,liver and its surrounding tissues are highly similar in gray scale,which causes liver segmentation results beyond the liver region.The imbalance of sample distribution and insufficient positioning information affect the effect of liver segmentation,so a unified and effective segmentation method has not yet been formed,which cannot be directly applied to clinical diagnosis and treatment.To address those problems,this paper puts forward corresponding improvement scheme.The innovation of this paper mainly has the following two points,summarized as follows:(1)A RV-DeepLab model which combines multi-level features and preserves mid-level features,is proposed to solve the problem of insufficient location information in the classical semantic segmentation DeepLab model.The proposed model uses the jump connection to integrate multi-level features and reintroduce the useful information of middle layer,balancing the two contradictory goals of location and classification in semantic segmentation.Compared with DeepLab,RV-DeepLab model has improved segmentation result on both 3Dircadb and Sliver07 datasets,with the score improved by 0.19 on Sliver07 dataset.(2)Aiming at the problem that the partial segmentation result of RV-DeepLab model is far beyond liver region,a DSL model is proposed to reduce the scope of semantic segmentation.Firstly,the model uses the online hard example mining algorithm to improve Faster R-CNN algorithm,solving the imbalance of sample distribution in liver detection.Improved Faster R-CNN can detect the general location in liver,narrowing the scope of subsequent semantic segmentation.Furthermore,the processed resulting image are input into RV-DeepLab to be segmented.Experimental results show that compared with RV-DeepLab model,DSL model has better segmentation effect on 3Dircadb and Sliver07 dataset,with the score increased by 5.75 on Sliver07 dataset.Meanwhile,DSL model is superior to other comparison methods.In conclusion,experiments and analysis verify that our RV-DeepLab and DSL model which can narrow the semantic segmentation range are effective.