【作者】 王涛；

【导师】 张永刚；

【作者基本信息】 吉林大学 ， 工程硕士（专业学位）， 2023， 硕士

【摘要】 核磁共振成像(Magnetic Resonance Imaging,MRI)越来越多的被临床医生用于评估和护理疑似脑损伤的新生儿。它提供了有关新生儿大脑的详细解剖信息,弥补了超声波在筛查和诊断脑部疾病方面的功能,能准确分割白质、灰质和脑脊液等脑部区域,对于新生儿早期大脑发育定量研究具有重要意义。特别是早产儿,由于其不正常的胎中发育时间,大脑往往区别于足月儿,此时通过脑部功能区域分割结果,可以协助医生更准确评估和诊断早产儿的脑部发育情况,同时给出准确的治疗方法,减少后遗症的概率。传统上,放射科医生一般通过手动标记来勾画出大脑结构,这是耗时且难以满足实际要求的。在过去十几年中,深度学习不断迭代发展,引起了医学领域的广泛关注。为了提高测量效率和准确性,使用深度学习技术实现医学图像自动分割成为了越来越多科研人员的追求。目前已经提出了许多用于自动图像分割的方法,但传统的卷积神经网络分割精度仍有进一步提高的空间。同时对于新生儿脑部区域分割,与成人大脑相比新生儿大脑的MRI图像显示出低信噪比、低组织对比度、部分体积效应以及白质之间的亮度反转,这对新生儿大脑结构的精确研究提出了严峻的挑战。Transformer在NLP中的巨大成功激发了将它用于其他任务的研究热情。Vision Transformer(VIT)作为一种纯Transformer体系结构首次被用于视觉任务,然而其存在高分辨率图像的兼容性问题,对于图像处理任务来说并不友好,提高了图像处理任务的硬件门槛。为了解决这一问题,Swin Transformer被提出,这可以说是原始VIT后最令人兴奋的研究之一。与VIT不同,其使用的移位窗口机制更高效且特征提取能力更强,由于其卓越性能表现,现在许多视觉模型体系结构中使用其作为模型的骨干,同样在医学图像分割领域也常有涉及。为了更精准的研究和评估新生儿脑部发育情况,本文开展了以下两个方面的工作:1.本文设计了一种全新的端到端医学图像分割框架,主要针对新生儿脑部MRI图像,包括新生儿脑部MRI图像预处理,训练,后处理等措施。预处理部分应用了N4偏置场校正以及平衡对比度增强等技术,提高输入图像的质量。训练部分,将Swin Transformers中的移位窗口自注意力模块提取出来加以改进,搭建编码器-解码器结构网络,配合全尺度跳跃链接以及深度监督机制,得到新生儿脑部多功能区域分割图像。后处理部分使用了最大连通域算法,提高分割精度。最后在d HCP新生儿脑部公共数据集上对该网络进行了训练和测试,并与先进的方法进行了定性和定量比较,实验证明本文提出的方法有效提高了分割精度,同时通过消融实验证明了本文提出改进的有效性。2.本文进一步设计了一个年龄预测框架,对于分割后的区域进行了进一步研究,根据新生儿脑部MRI来预测新生儿脑龄是否符合预期,更加直观的研究新生儿脑部发育情况。框架主要包含预处理以及网络训练两部分,预处理部分除了前一研究提出的预处理方法外,本研究还加入了脑颅骨去除,用以减少无关区域的影响。训练部分在前一研究基础上稍加改进,使其更加适合进行年龄预测。训练阶段最后同样在d HCP数据集上训练验证,与多个方法进行了定量比较,本文提出的方法在MAE以及R~2两项指标上均优于其他方法。更多还原

【Abstract】 Magnetic Resonance Imaging(MRI)is increasingly used by clinicians to evaluate and care for newborns with suspected brain injury.It provides detailed anatomical information about the neonatal brain,compensates for the function of ultrasound in screening and diagnosing brain disorders,and can accurately segment brain regions such as white matter,gray matter and cerebrospinal fluid,which is important for quantitative studies of early neonatal brain development.Especially in preterm infants,the brain is often different from that of full-term infants due to their abnormal intrauterine development time.At this time,the segmentation results of functional brain regions can assist physicians in more accurately assessing and diagnosing the brain development of preterm infants,as well as giving accurate treatment methods and reducing the probability of sequelae.Traditionally,radiologists typically sketch brain structures by manually labeling them,which is time-consuming and difficult to meet practical requirements.Over the past decade or so,deep learning has evolved iteratively and has attracted a lot of attention in the medical field.To improve measurement efficiency and accuracy,automatic segmentation of medical images using deep learning techniques has become a growing pursuit for researchers.Many methods have been proposed for automatic image segmentation,but there is still room for further improvement in the accuracy of traditional convolutional neural network segmentation.Meanwhile,for neonatal brain region segmentation,MRI images of neonatal brain show low signal-to-noise ratio,low tissue contrast,partial volume effect and brightness inversion between white matter compared with adult brain,which poses a serious challenge to the accurate study of neonatal brain structure.The great success of Transformer in NLP has inspired research to use it for other tasks.Vision Transformer(VIT)was first used as a pure Transformer architecture for vision tasks,however,it suffers from compatibility issues with high resolution images,which is not friendly for image processing tasks and raises the image processing hardware threshold for image processing tasks.To solve this problem,Swin Transformer was proposed,which is arguably one of the most exciting researches since the original VIT.Unlike VIT,it uses a shift window mechanism that is more efficient and capable of feature extraction,and is now used as the backbone of many vision model architectures due to its superior performance,and is also often involved in the field of medical image segmentation.In order to study and evaluate neonatal brain development more accurately,this paper carries out two aspects of work as follows.1.In this paper,a novel end-to-end medical image segmentation framework is designed,mainly for neonatal brain MRI images,including measures of neonatal brain MRI image pre-processing,training,and post-processing.The preprocessing part applies techniques such as N4 bias field correction and balanced contrast enhancement to improve the quality of the input images.In the training part,the shift window selfattentive module in Swin Transformers is extracted and improved to build an encoderdecoder structure network with full-scale jump link and deep supervision mechanism to obtain multifunctional region segmentation images of neonatal brain.The postprocessing part uses the maximum connected domain algorithm to improve the segmentation accuracy.Finally,the network is trained and tested on the d HCP neonatal brain public dataset,and compared qualitatively and quantitatively with state-of-the-art methods.The experiments demonstrate that the proposed method in this paper effectively improves the segmentation accuracy,and the effectiveness of the proposed improvement is demonstrated by ablation experiments.2.In this paper,we further designed an age prediction framework for the segmented region to predict the neonatal brain age as expected based on neonatal brain MRI,and to study the neonatal brain development more intuitively.In addition to the preprocessing method proposed in the previous study,this study also added brain cranial bone removal to reduce the influence of irrelevant regions.The training part is slightly improved from the previous study to make it more suitable for age prediction.The training phase was finally validated by training on the d HCP dataset as well,and a quantitative comparison was made with several methods,and the proposed method outperformed the other methods in both MAE and R2 metrics.更多还原