PET/CT Tumor Segmentation Based On Variational Methods

Positron Emission Tomography(PET)is a functional imaging method at the molecular level.It can describe the difference between the lesion and the normal tissue on metabolism and function.PET imaging can detect the early lesions effectively,so it has become an important tool for the early diagnosis and treatment of cancer.Accurate tumor segmentation from PET images is crucial in many radiation oncology applications.It can ensure the maximum amount of dose delivered to the cancer cells and potentially reduce the side effects to the surrounding healthy tissues and vital organs.However,the tumor edges in PET are usually blurred due to the partial volume effect(PVE),and the PET image has the characteristic of high noise and intensity inhomogeneity,which make it difficult to accurately delineate the tumor from a PET image.Computed Tomography(CT)is an anatomical imaging technology.CT images provide the tissue structure information of the human body with higher spatial resolution than PET.Combining the complementary information of PET and CT can improve the accuracy of tumor segmentation.This paper proposed a series of advanced tumor segmentation methods based on the features of PET and CT images.The fuzzy and high noise characteristics of PET images make it difficult to accurately locate the tumor edges.Taking into account that image restoration and tumor segmentation can promote each other,we integrated total variation(TV)semi-blind de-convolution with Mumford–Shah segmentation in a variational framework.The new designed variational model can complete PET restoration,tumor segmentation and blur kernel estimation of PET scanner at the same time.The proposed variational model contains a specially designed adaptive multiple regularizations that the TV regularization was used over tumor edges to preserve the edge information and the Tikhonov regularization was used in non-edge areas to preserve the smooth change of the metabolic uptake in a PET image.CT imaging has higher spatial resolution than PET.Fusing the CT image information can improve the accuracy of both PET restoration and tumor segmentation.We design a PET/CT multimodality tumor segmentation method integrated with PET restoration on the base of above simultaneous PET restoration and tumor segmentation model.This method combines the advantages of high contrast of PET and high spatial resolution of CT.It can automatically decide which modality should be more trustful when PET and CT disagreed to each other for localizing the tumor boundary.The structures of CT images are much complicated.The tumor intensity may be similar to that of the normal soft tissues in CT.The edges of normal tissues may interfere with the location of the tumor edges.Considering the excellent performance of deep learning method in handling complex problems,we further proposed a deep learning and variational method based PET/CT multimodality tumor segmentation method.Firstly,a 3D fully convolutional network(FCN)was constructed to extract the probability information from the CT images.Then,a fuzzy variational model was designed to incorporate the extracted probability information from CT and intensity information of PET to delineate the tumor more precisely.Although the deep learning method has shown obvious advantages in the field of image segmentation,it relies on a large number of labeled training samples.However,in medical image processing field,either acquiring that amount of training samples or obtaining their exact labels is generally infeasible.In order to more widely make use of the advantages of the deep learning network to deal with the medical image segmentation problems,we proposed a novel 3D FCN based unsupervised deep learning segmentation method.A task-specific(PET tumor segmentation)loss function was designed to guide the self-learning of the 3D FCN.Once the 3D FCN become mature by self-learning,it can perform the tumor segmentation task end-to-end.Finally,some open problems about PET and PET/CT image segmentation are summarized and future research directions are put forward.