Optimization and learning algorithms for medical image interpretation

总体背景和问题：每天获取和存储的大规模图像数据的压倒性增长为准确的预测模型、更好的指导决策以及新的见解和发现带来了前所未有的机会，并具有影响医疗保健、安全、社交媒体、机器人和自主系统、遥感和制造业等战略领域的巨大潜力。最先进的图像解释算法需要大量可靠的训练数据，即通过大量人力和专业知识构建的精确标记（注释）图像。这种全面的监督是广泛应用领域的主要障碍，例如医学图像分析。这需要将当前的算法提升到一个全新的自动化、可扩展性和准确性水平，利用大量可用的未标记图像和可能与图像（例如文本）相关的不确定（噪声）知识。目标：我们的研究重点是弱监督语义图像分割和分类的新数学模型和计算方法，这是图像解释系统中的两个关键问题。根据我们的专业知识，我们打算追求基于优化的配方，利用大规模的，大部分未标记的图像数据，这些数据具有重要和复杂的先验知识，这些知识在当前方法中被省略或过度简化。具体的技术目标是：(i)定义新的约束，这些约束体现了领域知识，从而减轻了数据注释的缺乏/不确定性；（ii）设计新颖的，基于近似的策略来优化随之而来的难题；（iii）评估和验证我们对医学图像的调查，因为它们的高度可变性和复杂性，它们给机器学习算法带来的挑战（例如，缺乏注释），以及它们对推进医疗保健实践/研究的巨大承诺。虽然这项研究可以服务于广泛的临床应用，但我们打算优先考虑对经济和社会有重大影响的两个领域：神经病学和肿瘤学。意义：通过整合领域知识和大规模数据的公式，以及先进的优化专业知识，本研究有望提供具有国际竞争力的算法（在自动化，精度，速度和鲁棒性方面）。其范围远远超出了计算机视觉和医学成像的应用，在一般的、广泛的弱监督学习主题和各种应用学科中具有潜在的影响。仅在医学成像方面，考虑到丰富的领域知识（例如，解剖学和放射学文本报告），潜力是巨大的。在这个应用领域，强大的弱监督算法有望影响医疗保健研究（例如，对复杂疾病的理解）和实践（例如，疾病的早期检测、诊断、监测、治疗和随访）。

General context and problematic: The overwhelming growth of large-scale image data acquired and stored everyday brings unprecedented opportunities for accurate predictive models, better-guided decisions and new insights and discoveries, with the enormous potential to impact strategic areas such as health care, security, social media, robotics and autonomous systems, remote sensing and manufacturing. State-of-the-art image interpretation algorithms require very large amounts of reliable training data, i.e., accurately labeled (annotated) images built with extensive human labour and expertise. Such a comprehensive supervision is a major impediment in a breadth of application areas, e.g., medical image analysis. This calls for bringing current algorithms to a whole new level of automation, scalability and accuracy, leveraging the available large-scale amounts of unlabeled images and the uncertain (noisy) knowledge that might be associated with the images (e.g., text). Objectives: Our research program focuses on novel mathematical models and computational methods for weakly supervised semantic image segmentation and categorization, the two key problems in image interpretation systems. Following on our expertise, we intend to pursue optimization-based formulations, which leverage large-scale, mostly unlabeled image data with important and complex prior knowledge that has been either omitted or oversimplified in current methods. Specific technical objectives are: (i) defining novel constraints, which embody domain knowledge, thereby mitigating the lack/uncertainty of data annotations; (ii) designing novel, approximation-based strategies for optimizing the ensuing difficult problems; and (iii) evaluating and validating our investigations on medical images, for their high variability and complexity, the challenges they bring to machine learning algorithms (e.g., the lack of annotations), and their great promises for advancing health-care practices/research. While this research can serve a breadth of clinical applications, we intend to prioritize two domains of high impact on the economy and society: neurology and oncology. Significance: With formulations integrating domain knowledge and large-scale data, as well as advanced optimization expertise, this research promises to deliver internationally competitive algorithms (in terms of automation, precision, speed and robustness). The scope goes far beyond computer-vision and medical-imaging applications, with a potential impact in the general, wide-interest subject of weakly supervised learning, and in various application disciplines. In medical imaging alone, the potential is huge given the rich domain knowledge (e.g., the anatomy and radiology text reports). In this application domain, powerful weakly supervised algorithms promise to impact health care research (e.g., the understanding of complex diseases) and practices (e.g., disease early detection, diagnosis, monitoring, treatment and follow-up).