New methods for automatic quality assessments of medical image registration

图像配准是计算机视觉和图像处理中最常见的任务之一，它是将两幅图像对齐，使其对应的特征在空间上匹配的过程。在放射诊断、手术计划和医学图像分析的应用中，医学扫描之间的准确配准的要求是普遍存在的，并且对于患者的安全、治疗结果和基于图像的疾病机制研究的可靠性至关重要。通常使用自动配准算法。然而，可能导致不利甚至致命后果的次优结果仍然可能发生，而质量控制主要依赖于主观目视检查，这主要是由于缺乏地面实况。因此，有效的自动技术来评估和可视化配准误差和不确定性是非常有价值的，但仍然是探索不足。目前，医学图像配准质量评价主要存在三个方面的问题。首先，在缺乏基本事实的情况下，通常难以评估配准的质量（例如，匹配解剖学标志和图像之间的结构分割），但是这些产生起来是耗时的，并且受到与评价者相关的可变性的影响。其次，自动技术在配准质量评估的模态间和对比度间对齐是罕见的。最后，需要专门的方法来量化和可视化基于深度学习的图像配准算法的配准质量，以提高其可解释性。拟议的研究计划将解决这些挑战，并开发新的技术，以提供有效和强大的评估医学图像配准质量。更具体地说，我们将提出自动解剖标志识别的新方法，设计基于学习的技术来直接预测配准误差，并建立新的框架，可以量化和可视化基于深度学习的配准算法的配准不确定性。该研究将满足对稳健方法的迫切需求，以量化和可视化图像配准质量，用于开发和验证新的图像配准算法。在患者护理和疾病分析的应用中，所得到的方法有望为最终用户提供自动图像配准软件结果的快速反馈，以确保疾病诊断的准确性，图像引导手术中的患者安全性以及涉及放射扫描的大数据分析的可靠性。

As one of the most common tasks in computer vision and image processing, image registration is the procedure of aligning two pictures so that their corresponding features can be spatially matched. In the applications of radiological diagnosis, surgical planning, and medical image analysis, the requirement of accurate registration between medical scans is ubiquitous and crucial for the patient's safety, treatment outcomes, and reliability of image-based research into the mechanisms of diseases. Automatic registration algorithms are commonly used. However, sub-optimal results that may lead to adverse and even fatal consequences can still occur while quality control primarily relies on subjective visual inspection largely due to the absence of ground truths. Therefore, effective automatic techniques to assess and visualize registration errors and uncertainties are highly valuable, but are still under-explored. Currently, there are three main challenges in the assessment of medical image registration quality. First, the quality of a registration is often difficult to evaluate in the absence of ground truths (e.g., matching anatomical landmarks and structural segmentation between images), but these are time-consuming to produce and subject to rater-dependent variability. Second, automatic techniques in registration quality assessment for inter-modal and inter-contrast alignment are rare. Lastly, specialized methods are needed to quantify and visualize registration quality for deep-learning-based image registration algorithms to improve their interpretability. The proposed research program will tackle these challenges and develop novel techniques to provide efficient and robust assessment of medical image registration quality. More specifically, we will propose new methods for automatic anatomical landmark identification, devise learning-based techniques to directly predict registration errors, and establish novel frameworks that can quantify and visualize registration uncertainties for deep-learning-based registration algorithms. The proposed research will meet the urgent needs for robust methods to quantify and visualize image registration quality for the development and validation of new image registration algorithms. In the applications of patient care and disease analyses, the resulting methods are expected to provide fast feedback to the end users for the results from automatic image registration software to ensure the accuracy of disease diagnoses, patient safety in image-guided surgery, and reliability of big data analyses involving radiological scans.