Medical Image Segmentation Based On Deep Learning

In the field of medical images,automatic segmentation from medical images is an important step for computer-aided cancer diagnosis and treatment.Recently,with the rapid development of deep learning,image segmentation techniques based on convolutional neural networks have gradually matured.However,due to the particularity of the task and the limitations of the deep learning method,the segmentation algorithm still has a lot of room for improvement.This thesis proposed a series of advanced segmentation methods based on the features of image and deep learning.PET,CT and MRI scanner is widely used in the clinic.Positron Emission Tomography(PET)is a functional imaging method at the molecular level.PET imaging can detect the early lesions effectively,so it has become an important tool for the early diagnosis and treatment of cancer.However,the lesion edges in PET are usually blurred because of the partial volume effect(PVE),and the PET imaging has the properties of intensity inhomogeneity and high noise.Computed Tomography(CT)is an anatomical imaging technology,which provide the tissue anatomical information of the human body with higher spatial resolution than PET.However,it is similar between lesion and surrounding tissues.Therefore,it is very difficult to perform segmentation task using single modality.In this study,we proposed a co-segmentation method based on a 3D fully convolutional neural network(FCN),which is capable of taking account of the advantages of both PET and CT into the same utility and completes the pixel-level segmentation task.It is worth mentioning that the proposed co-segmentation method is a general network architecture which can be transferred to other modalities or tasks.Magnetic Resonance Imaging(MRI)is often used for multi-organ segmentation tasks due to its high resolution.Because of the diversity of organs,their features are also different,mainly including shape,grayscale and texture feature,which will bring about the problem of multi scale in semantic segmentation.In order to improve the segmentation accuracy,this thesis proposed a more optimized method for multi organ segmentation in MRI image.The method can be integrated into other segmentation networks in a plug-and-play manner,it is very easy to use.