Liver Tumor Segmentation Based On Multi-parameter Spectral CT Images

Hepatocellular carcinoma(Hepatocellular Carcinoma,HCC)is one of the most lethal cancers in the world.In the course of HCC treatment,accurate segmentation of the lesion area based on medical images is very important for the delineation of tumor boundary and the measurement of tumor volume.Conventional CT has high resolution and has become a common technique for qualitative and localizing diagnosis of hepatocellular carcinoma(HCC).As a new type of CT scanning technology,spectral CT imaging can transform mixed energy imaging into single energy imaging while preserving the conventional CT function,and provide a variety of imaging information,such as spectral curves,diagrams of base material decomposition and the effective atomic number other than traditional imaging.Compared with conventional CT scanning,spectral CT can improve soft tissue resolution and provide a new method for the diagnosis of hepatocellular carcinoma(HCC).At the same time,spectral CT images with various parameters reflect tumor information from a variety of perspectives,and so that it could discriminate the tumor more accurately.Clinically,it is complicated and time-consuming to complete HCC tumor segmentation manually by doctors,and it is easy to cause subjective differences due to the individual experience of doctors in diagnosis.In recent years,machine learning,especially deep learning,has made a great breakthrough in the field of medical image processing.Deep learning algorithm can independently learn tumor features,mining potential high-dimensional semantic information,and achieve the purpose of automatic tumor recognition and segmentation.Therefore,we first apply deep learning to HCC tumor segmentation based on multi-parameter spectral CT images.To be accurate,objective and fast,we construct a deep network of multi-input mode by using the result of multi-energy spectral CT image segmentation.In order to learn from multi-parameter inputs to improve HCC tumor segmentation,we proposed two segmentation ideas based on late fusion network and multi-path densenet.First,according to the experience of clinicians,we used monochromatic images based on 45,55 and 65 keV,arterial phase iodine map,and customized U-Shape architecture to segment the lesions respectively as the control group for our multi-parameter segmentation results.We collected data from 18 patients and completed 6-fold cross-validation.Secondly,we make 45 keV and 65 keV pairing as two inputs of the late fusion network(Late Fusion Network,LFN),in order to combine the features from multiple images and improve the segmentation results.Similarly,we used spectral CT data from 18 patients to complete 6-fold cross-validation.Statistical analysis of the segmentation results shows that the segmentation of multiple images is much better than that based on single parameter spectral CT images.Finally,in order to overcome the deficiency of using multi-parameter image information of late fusion network structure both in principle and implement,we propose a segmentation scheme based on multi-path dense connection Network(Multi-path DenseNet,MDN).At the same time,to investigate the degree of usability of multi-parameter features,we designed a feature reuse experiment.The results show that the ability for segmentation for MDN was better than LFN.Compared with single parameter segmentation,MDN can effectively utilize the multi-parameter information of spectral CT.The above results show that the proposed deep learning method achieved the information fusion of multi-parameter spectral CT images,and the segmentation performance is better than that based on single parameter and U-Shape network architecture.It shows the good prospect of HCC tumor segmentation based on deep learning and spectral CT multi-parameter image.However,the proposed segmentation method is based on two-dimensional images and fails to make full use of the spatial correlation of three-dimensional data.In addition,our training and testing process was based on the images of tumor-containing,and the segmentation performance on non-tumor-free spectral CT images has not been verified.In the next step,we consider to use three-dimensional MDN architecture for HCC tumor segmentation.Meanwhile,train and test on all spectral CT images(including or without HCC tumor foci)to increase the robustness of the algorithm with good exception on providing support for the clinical diagnosis and treatment of HCC.