Interdependent 3D Segmentation of Abdominal Organs in MRI Data Using Multiple Level Set Methods in Organ-Specific Probability Maps

医学图像数据中的腹部器官分割为提高诊断和流行病学标准提供了多种可能性，因此对加速现代医学研究特别有意义。这个项目提案是为流行病学研究中的3D分割而设计的，在该研究中，产生了超过3500个来自志愿者的磁共振(MR)数据集，包括上腹部部分。对于需要(完全)自动分割支持的如此大量的MR体积数据，手动分割是非常复杂和耗时的。尽管在MR体数据中进行全自动3D分割是非常具有挑战性的，但本项目的特别动机是开发一个具有很高价值的全自动分割框架，可以同时分割多个腹部器官，多器官分割方法采用联合勾画策略同时分割相邻器官，因此可以比单器官分割方法更有效地减少溢出。然而，现有的多器官分割方法都是为CT数据分割而设计的，存在严重的缺陷。我们提出了一种新颖的、自动的磁共振体积数据分割方法，该方法将利用器官特定特征的先验知识，并将其整合到分层组织的多步3D分割框架中的扩展算法中。在先前的工作中，我们证明了概率图(PM)有效地支持单个目标器官的分割，并将这一成功的概念扩展到使用与勾画相关的器官特征的多器官分割。由于一些腹部器官表现出局部不同的MR强度分布，我们将考虑器官内部的位置。结合相关器官特征的高性能机器学习技术将被应用于生成有意义的PM。我们将考虑流行病学研究的特殊要求，如组织类型适应的器官分割(例如，对于脂肪肝)和由于外科手术而可能被移除的器官的检测。多水平集方法将被应用于器官特定概率图，根据相互依赖的分割策略同时分割多个目标器官。每个零能级集描绘一个器官，并由最小化的器官特定能量项控制。将使用新的3D形状描述符来确定将在先前的基于形状的水平集分割部分中使用的准确的器官形状分布。我们的模块化分割框架将按层次组织，第一层分割肝脏、肾实质和脾，第二层分割胆囊和胰腺。然而，它可以灵活地扩展用于未来的使用，也将是适用的，因此对临床常规的分割具有很高的价值。

Segmentation of abdominal organs in medical image data offers miscellaneous possibilities to advance diagnostic and epidemiological standards and is therefore of particular interest to accelerate modern medical research. This project proposal is designed for 3D segmentation in an epidemiological study, in which more than 3500 magnetic resonance (MR) datasets from volunteers are produced including the upper abdominal body part. Manual segmentation is very elaborate and time-consuming for such high amount of MR volume data requiring (fully) automatic segmentation support. Although fully automatic 3D segmentation in MR volume data is very challenging, the particular motivation of this project is to develop a highly valuable fully automatic segmentation framework that can segment several abdominal organs simultaneously.Multi-organ segmentation approaches segment adjacent organs simultaneously in a combined delineation strategy and can therefore reduce overspills more effectively than single-organ segmentation methods. However, existing multi-organ approaches are designed for CT data segmentation and have severe drawbacks. We propose novel, automatic methods for MR volume data segmentation that will use prior knowledge of organ-specific features and incorporate them in extended algorithms in a hierarchically organized multi-stepped 3D segmentation framework.In previous work we showed that probability maps (PMs) support segmentation of single target organs efficiently and we will extend this successful concept for multi-organ segmentation using delineation-relevant organ features. Since some abdominal organs show locally varying MR intensity distributions, we will take inner-organ locations into account. Highly performant machine learning techniques incorporating relevant organ features will be applied to generate meaningful PMs. We will consider particular requirements of epidemiological studies like tissue type adapted organ segmentation (e.g. for fatty livers) and detection of potentially removed organs due to surgical operations.A multiple level set method will be applied in organ-specific probability maps to segment several target organs simultaneously according to an interdependent segmentation strategy. Each zero level set delineates an organ and is controlled by minimized organ-specific energy terms. Novel 3D shape descriptors will be used to determine exact organ shape distributions that will be used in the prior shape-based level set segmentation part. Our modularized segmentation framework will be hierarchically organized for liver, renal parenchyma and spleen segmentation in the first level and gallbladder and pancreas segmentation in the second level. Though, it can be flexibly extended for future use and will also be applicable and therefor highly valuable for segmentation in clinical routines.