Graph-Based Medical Image Segmentation in 3D and 4D

描述（由申请人提供）：这是我们第二期项目的竞争性延续。在成功实现所有目标后，我们的优化多表面和/或多目标n-D生物医学图像分割框架得到了进一步的扩展和验证，并在临床和翻译图像分析任务中展示了其实际效用。这个第三阶段的建议将发展几个重要的扩展，以解决当前框架的局限性，并特别关注该方法对转译和常规医疗保健任务的适用性。将开发新的方法来同时分割相互作用的区域和表面，从分割示例中自动设计成本函数，并通过允许有限和高效的专家输入来指导图像分割过程，克服常规诊断质量图像中自动化技术的失败。我们假设，先进的基于图的图像分割算法融合了机器学习衍生的分割参数和特定于图像的专家指导，将显著提高在不同应用领域常规获取的复杂诊断质量医学图像的定量分析性能。我们建议：1)开发3D、4D和n-D方法来同时分割相互作用的区域（物体）和表面。2)开发数据驱动的成本函数自动化设计方法，用于基于表面、基于区域、基于表面和基于区域的图搜索图像分割。3)发展“刚够互动”（Just-Enough-Interaction， JEI）方法，实现高效的“实时”医学图像分割，实现医学图像定量分析的强大临床适用性。4)评估所有已开发方法在转化研究环境中的表现；确定定量医学图像分析和放射肿瘤学治疗计划工作流程的性能。因此，我们的项目将实现常规量化，从而实现个性化护理。

DESCRIPTION (provided by applicant): This is a competitive continuation of our Phase-II project. After successfully fulfilling all of its aims, our framework for optimal multi-surface andor multi-object n-D biomedical image segmentation was further extended, validated, and its practical utility demonstrated in clinical and translational image analysis tasks. This Phase-III proposal will develop several important extensions addressing identified limitations of the current framework and specifically focusing on applicability of the methodology to translational and routine healthcare tasks. Novel methods will be developed for simultaneous segmentation of mutually interacting regions and surfaces, automated design of cost functions from segmentation examples, and overcoming failures of automated techniques in routine diagnostic quality images by allowing limited and highly efficient expert input to guide the image segmentation processes. We hypothesize that advanced graph-based image segmentation algorithms merging machine- learning-derived segmentation parameters and image-specific expert guidance will significantly increase quantitative analysis performance in routinely acquired complex diagnostic-quality medical images across diverse application areas. We propose to: 1) Develop 3D, 4D, and generally n-D approaches for simultaneous segmentation of mutually interacting regions (objects) and surfaces. 2) Develop methods for data-driven automated design of cost functions used for surface-based, region-based, and surface-and-region-based graph search image segmentation. 3) Develop "Just-Enough-Interaction" (JEI) approaches for efficient "real-time" medical image segmentation, thus achieving robust clinical applicability of quantitative medical image analysis. 4) Assess performance of all developed methods in translational research settings; determine performance in quantitative medical image analysis and radiation oncology treatment planning workflow. As a result, our project will enable routine quantification and therefore personalized care.