Research On Fetal Magnetic Resonance Image Based On Deep Learning

Recently,with the rapid development of magnetic resonance imaging(MRI),fetal MRI has gradually become an important tool for clinical early diagnosis of fetal abnormalities and brain development analysis.Fetal MRI image quality assessment and brain segmentation are the basis of 3D reconstruction and quantitative analysis of the fetal brain.Fetal brain age prediction based on fetal MRI images also plays a vital role in the diagnosis of fetal early disease and brain development analysis.At present,there are still many challenges in the research of deep learning algorithms based on fetal MRI images:(1)Data collection is difficult and the amount of data is scarce?(2)There are many problems in the quality of data images,such as noise and blur?(3)Shape,structure and volume of fetal brain are changeable?(4)Interference of the mother’s organs around the fetus?(5)Differences in the brains of different individuals.Based on the above challenges,the main research content and contributions of this article are as follows:For image quiality assessment tasks and brain segmentation tasks,we propose a multi-task deep learning framework to perform image quality assessment and brain segmentation simul-taneously.Taking into account the characteristics of fetal MRI images,a brain detector is first designed to locate the brain area.Then,a deformable convolution that can adaptively adjust the receptive field is introduced to deal with the problem of changing brain shape and structure and random orientation.Finally,two task-specific modules are used to perform quality assessment and brain segmentation respectively.For model performance,we further propose a multi-step training strategy.For the age prediction task,we propose a network framework based on deformable convo-lution to predict the age of the fetal brain.Taking into account the scarcity of fetal MRI data,the label distribution learning algorithm used to solve the small sample problem is introduced,and the age label probability distribution information is integrated into the end-to-end network framework.In order to further utilize the multi-view image information obtained during fetal scanning,a multi-branch network structure is designed to fuse these complementary multi-view images.In summary,this paper studies image quality assessment,brain segmentation and age pre-diction based on fetal MRI images.We ntroduce the multi-stage network process,multi-task joint learning iframework,deformable convolution,label distribution learning algorithm,etc.to solve the difficulties and challenges brought by fetal MRI data.This paper conducts corre-sponding ablation experiments and comparison experiments with other algorithms on the fetal MRI data set collected by ourself.The experimental results prove that our method has excellent performance for image quality evaluation,brain segmentation and age prediction.