基于三维卷积神经网络的肝脏MRI图像自动分割方法的研究

The segmentation of the liver including tumor and hepatic vessels is a challenging task in medical image processing. However, current methods mainly focus on CT scans and the segmentation techniques for MRI images remain limited. Delineating the liver including tumor and hepatic vessels is extremely challenging from multi-phase contrast-enhanced MRI images. For large medical data, variational based method, a semi-automatic approach, is limited in the clinical application, while deep learning has an advantage in extracting and selecting features automatically. Based on the medical background of accurate liver diseases diagnosis and surgery planning, this project is to conduct the research of fully automatic segmentation modeling based on the combination of deep learning and variational method. Under the framework of from coarse to fine, on three main scientific problems, i.e., the automatic segmentation of multi-object, the co-segmentation of multi-phase MRI images and automatic segmentation based on small labeled data, this project is to model some automatic and accurate segmentation, through the combination of deep model, variational approach, complex constraint and prior shape constraint, to overcome complex backgrounds, inhomogeneous appearance and weak boundary in MRI images and achieve the rapid and accurate 3D MRI image automatic segmentation.

肝脏、肿瘤及血管的自动分割是当前医学图像处理中一个难点。当前的分割方法主要集中在腹部CT图像上，而在MRI图像上的研究工作有限。从多期增强MRI图像上自动分割肝脏、肿瘤以及血管也更具挑战性。面临医学大数据，半自动的变分分割法的实际临床应用有限，但深度学习模型具有自动提取与筛选特征的显著优势。本项目立足于肝脏疾病精确诊断与手术方案规划的医学应用大背景，开展基于深度学习与变分法的MRI图像全自动分割方法的建模与分析的研究。在从粗到细的框架下，围绕肝脏MRI图像多目标自动分割、多期MRI数据协同分割以及小样本训练集下肝脏MRI图像自动分割三个主要问题，借助于深度学习、变分能量法以及解剖约束关系和先验形状约束建立一系列全自动精准分割模型，克服肝脏MRI图像背景复杂、前景表观不均匀、边界模糊等难点，实现三维MRI图像的快速、自动、精准分割。

肝脏分割是临床肝脏手术方案制定的基础，然而手动分割耗时耗力，为此本项目基于三维卷积神经网络的方法，在小样本数据集上开展了肝脏自动分割的深度建模。首先借助于3D DenseNet构造肝脏形状自编码器提取肝脏先验形状；然后基于DenseNet、FCN和U-Net构造肝脏自动分割3D深度卷积神经网络结构，并通过融合自适应的交叉熵损失函数、边界正则项和形状约束项提出了一种混合损失函数。通过公共数据集上的数值实验，验证了该网络在混合损失函数的指导下，即使在小样本训练集上，也能准确提取分割结果，无需耗时的后处理就能准确地提取分割结果。