Computational optimization techniques for medical image analysis

我建议为医学图像分析（MIA）开发高度自动化、鲁棒性和准确性的计算技术。新的图像处理、数据分析、人工智能和人工生命技术将在医学背景下进行研究和验证。医学成像提供了具有挑战性的算法问题，并具有改善健康的潜力。挑战主要是由于大的解剖形状可变性（例如肱二头肌槽），非刚性解剖（例如心肌），嘈杂的医学图像（例如MRI， CT或SPECT），数据多样性（例如2D彩色显微镜，3D MRI，时变PET或扩散张量（DT） MRI），以及与其他领域相比，医学对鲁棒性和准确性的更高要求。优化问题是MIA中经常遇到的问题。我建议的研究包括：1)调查现有优化技术的局限性和优势，以理解医学图像中的视觉数据；2)以适合启发式和领域知识的直观编码的新方式制定MIA问题；3)扩展优化技术以获得更好的收敛性和最优性。我将重点讨论以下问题：(I)医学图像分割（识别图像中结构和器官的边界，例如用于后续量化）；（ii）医学图像配准（图像的空间对齐，例如用于构建变分地图集或融合多模态图像）；（iii）结构形状对应（配对形状上的同源点，例如用于形状分类和统计分析）。对于分割，我将标量图像分割算法扩展到DTMRI数据，并提出替代的知识驱动优化公式。对于图像配准，我将优化配准策略，而不是使用低级指标作为目标函数。我还将开发形状对应的替代优化方法，并提出形状对应和图像配准的统一框架。

I propose to develop highly-automated, robust and accurate computational techniques for medical image analysis (MIA). Novel image processing, data analysis, artificial intelligence, and artificial life techniques will be investigated and validated in a medical context.          Medical imaging provides challenging algorithmic problems and has a potential for improving health. The challenges are primarily due to the large anatomical shape variability (e.g. humeral bicipital groove), non-rigid anatomy (e.g. myocardium), noisy medical images (e.g. MRI, CT, or SPECT), data variety (e.g. 2D colour microscopy, 3D MRI, time-varying PET, or diffusion tensor (DT) MRI), and higher demand for robustness and accuracy in medicine compared to other areas.          Optimization problems are often encountered in MIA. My proposed research involves: 1) investigating the limitations and strengths of existing optimization techniques for understanding the visual data in medical images; 2) formulating MIA problems in new ways amenable to intuitive encoding of heuristics and domain knowledge; and 3) extending optimization techniques to obtain improved convergence and optimality. I will focus on the problems of: (i) medical image segmentation (identifying the boundaries of structures and organs in images, e.g. for subsequent quantification); (ii) medical image registration (spatial alignment of images, e.g. for building variational atlases or fusing multi-modal images); and (iii) structural shape correspondence (pairing homologous points on shapes, e.g. for shape classification and statistical analysis). For segmentation, I will extend scalar image segmentation algorithms to DTMRI data and propose alternative knowledge-driven optimization formulations. For image registration, I will optimize registration strategies rather than using low-level metrics as objective functions. I will also develop alternative optimization approaches for shape correspondence and propose a unified framework for shape correspondence and image registration.