Medical Image Segmentation Based On Fusing Deep Learning And Anatomical Prior

Medical image segmentation is a critical step in Computer-Assisted Diagnosis(CAD)and the basis for further analysis of it.The medical images,such as CT and MRI,are gray-scale images with low-contrast and some neighboring organs in them have similar intensities and textures.Thus,accurate medical image segmentation is challenging.Currently,deep learning techniques achieves great success in numerous computer vision tasks including object detection,image classification,image segmentation and so on.Some of these deep learning methods are also widely applied to medical image research.However,for medical image segmentation tasks,further research can be made from the following two aspects.First,human organs usually present similar shapes and positions in medical images,which can be called the anatomical prior of the medical images.Therefore,we can consider the anatomical prior of the medical images and combine deep learning methods to improve the performance of organ segmentation.Second,the training of deep learning models requires a large amount of annotated data,however,the annotated datasets for medical image segmentation are usually small.Therefore,a semi-supervised learning method can be used,which trains the segmentation model using both annotation and unannotated data,to reduce the pressure on medical image annotation.According to the above analysis,this paper focuses on the organ segmentation task in medical images,researches on fusing deep learning and anatomical prior knowledge,and effectively uses anatomical prior knowledge of medical images to improve the performance of organ segmentation.At the same time,aimed at the lack of annotated medical images,two semi-supervised learning models using unannotated data and anatomical prior knowledge are proposed.The main contributions of this paper are listed as follows:1)Propose a medical image segmentation method based on anatomical prior.Firstly,the method incorporates medical image anatomical prior into the training of deep learning model through a DAP(Deep Atlas Prior)loss function.Secondly,regarding the DAP loss as a prior loss,and combining it with conventional likelihood loss,the adaptive Bayesian loss function is proposed,like the Bayesian framework,which consists of a prior and a likelihood.The adaptive Bayesian loss dynamically adjusts the ratio of DAP loss to likelihood loss during training.On a public liver dataset(from the ISBI LiTS 2017 Challenge)and a private spleen dataset(from a hospital),a series of experiments are performed comparing with different loss functions and the state-of-the-art segmentation models to verify the significance and generalization of the proposed methods.In the experiments,the Dice coefficients of liver and spleen segmentation reaches 96.05% and 95.77%,respectively.2)Propose a semi-supervised segmentation method based on adversarial learning.Aimed at the lack of annotated medical images,this method filters reliable pixels from unannotated data by the discrimination network in it for segmentation network training.At the same time,this framework uses the adaptive Bayesian loss function to fuse the anatomical prior features of medical images.Experiments are performed on the liver and spleen datasets.In the experiments,the training dataset is divided into annotated and unannotated data by different ratio.The experimental results show that the semi-supervised segmentation model uses unannotated data and anatomical prior features to improve the results of medical image segmentation effectively with different ratio,and the segmentation performance exceeds that of other semi-supervised models.(3)Propose a semi-supervised medical image segmentation method based on prior similarity.This method calculates the similarity from the probabilistic atlas and the segmentation prediction results as the confidence basis,and pick up high-confidence pixels from unannotated data to train the segmentation model.Experiments are performed on the liver and spleen datasets,and two semi-supervised medical image segmentation models based on adversarial learning and prior similarity are compared.Experimental results verify that the segmentation performance of latter surpasses the former.