Learning Algorithms for Predictive Modeling in Biomedical Computing: Methods and Applications

计算机辅助干预（CAI）是指将来自众多生物传感器的信息与计算算法结合起来以帮助决策的系统。CAI系统是现代医学的基石，在过去的几十年里在放射学和机器人手术等领域发挥了重要作用。从各种生物传感器和工具（例如成像模式）生成的大量CAI数据得到了计算方法（特别是用于数据分析的机器学习）前所未有的进步的补充。因此，对生物医学数据的理解受到了根本性的影响，主要是以积极的方式，预计这一趋势将继续增长。然而，在实践中，丰富的机器学习解决方案与它们在决策中的应用形成了鲜明的对比。实现这一点的重大挑战是来自各种模式和人群的数据的噪声和高度异质性，以及注释的有限可用性和不准确性。 该提案的总体目标是设计下一代学习算法和关键决策方法，用于可操作的优化CAI，以应对该领域开发技术的挑战。我提出了创新的方法，即：i）从特定于模态的属性中分离出数据的特定于任务的信息属性。然后，可以使用特定于任务的属性来确保从一个数据域（例如，成像模态，诸如原始US）学习的知识可以转移到其他域（例如，B模式US）。为了融合数据的多种形式，设计了灵活有效的基于深度学习的方法，该方法消除了配准期间的优化并提供不确定性估计; ii）涉及通过数据点之间的关联的无监督发现从具有不精确注释的数据中学习的判别和生成方法，以及在有限的可用黄金标准注释的背景下使用这种关联进行决策; iii）使用同时的理论引导和数据驱动的学习为组织分类提供决策支持。基于理论的模拟数据允许收敛到解决方案，实验数据有助于通过无监督对抗方法最大限度地减少前一步的残余误差。 拟议的研究计划将提供培训3博士，3硕士和5 UG HQP。将为HQP提供最先进的跨学科培训环境，遵循公平，多样性和包容性的原则。学员将获得广泛的技能，包括图像处理，机器学习，决策和软件开发。他们将学习将他们的计算知识转化为具有实际意义的高影响力问题。他们的培训将对公共和私营部门在机器学习和生物医学方面日益增长的需求具有相当大的价值。

Computer Aided Interventions (CAI) refer to systems that incorporate information from a multitude of biosensors, with computational algorithms to assist in decision making. CAI systems are a cornerstone of modern medicine and have played a significant role for the past few decades in areas such as radiology and robotic surgery. The vast amount of data generated from various biosensors and tools (e.g. imaging modalities) for CAI are complemented by unprecedented advances in computational approaches in particular machine learning for analysis of data. Understanding biomedical data has therefore been impacted fundamentally, mostly in a positive manner, with this trend expected to grow. However, in practice, the abundance of machine learning solutions stands in stark contrast to their uptake in decision making with practical impact. Of the significant challenges for realizing this are the noisy and highly heterogenous nature of the data from various modalities and populations, and the limited availability and inaccurate nature of annotations. The overarching goal of this proposal is to design the next generation of learning algorithms and critical decision-making approaches for actionable, optimized CAI that address the challenges to the uptake of the developed techniques in this domain. I propose innovative methods that: i) disentangle informative task-specific attributes of data from modality-specific attributes. Task-specific attributes can then be used to ensure knowledge learned from one domain of data (e.g. imaging modality such as raw US) is transferable to other domains (e.g. B-mode US). To fuse multiple modalities of data, flexible and efficient deep learning-based approaches that eliminate optimization during registration and provide uncertainty estimates are devised; ii) involve discriminative and generative approaches to learn from data with imprecise annotations through unsupervised discovery of associations between data points, and decision making using such associations in the context of limited available gold-standard annotations; iii) provide decision support for tissue classification using simultaneous theory-guided and data-driven learning. Theory-based, simulated data allow convergence to a solution and experimental data help minimize the residual error from the previous step, through unsupervised adversarial methods. The proposed research program will provide training for 3 PhD, 3 MSc and 5UG HQP. State of the art interdisciplinary training environment for HQP that follows the principals of Equity, Diversity and Inclusion will be provided. Trainees will acquire a wide spectrum of skills including image processing, machine learning, decision making and software development. They will learn to translate their knowledge of computing to high impact problems with practical implications. Their training will be of considerable value to a growing demand in both public and private sectors in machine learning and biomedicine.