Not-So-Supervised Meta-Learning for Medical Image Analysis

在医学成像应用中应用现代深度学习方法已经显示出自动化的潜力，并进一步改善了广泛的临床应用，如解剖识别、定位和病理量化。然而，由于需要大量的专家标记训练数据，与标签相关的不确定性以及解释预测的困难阻碍了这些方法的临床应用。本项目研究了一种元学习策略，该策略可以将弱监督、半监督和无监督技术的好处推广到前列腺癌和肺部疾病诊断的成像应用中，克服这些局限性。2)目的和目标-具体目标是：该项目旨在开发新的元学习方法，以解决基于深度学习的医学图像分析中的泛化问题，特别是在元学习中结合弱监督、半监督和无监督算法，以最大限度地减少对主观临床标签的依赖。建议的研究将集中在学习潜在的监督信号以及任务和元优化阶段的内在数据结构。本项目的核心假设是，通过结合学习策略的两种互补视角，医学图像分析应用将受益最大：虽然元学习利用其在任务之间学习“先验”知识的潜力，但不受监督的算法学习“内循环数据”之间的“内部”数据知识。不仅要克服标记数据的限制，一个更有趣的方向是量化表示，学会改进小数据训练，发现预测的推理或将这些预测与临床可解释的测量联系起来。3)研究方法的新颖性本项目计划研究弱监督、半监督和无监督算法在元学习中的作用，探索两个嵌套循环中潜在的监督信号。在内环中，将使用从无监督方法中学习到的“伪标签”来加强监督，提高目标任务的性能。在外环中，将使用度量学习来研究任务相似度，提高对目标任务的泛化。此外，相关分析将用于量化数据结构和与已知可解释知识的相似性，如临床测量和标签质量。这有助于测量模型和数据的不确定性，并解释机器预测，如放射科医生的注释和分数。4)与EPSRC的战略和研究领域保持一致。该项目与EPSRC的“医疗技术”主题和“人工智能技术”研究领域保持一致。通过这个项目，我们的目标是解决“物理干预的前沿”挑战，这是医疗技术主题战略的一部分。该项目还将通过基于医学图像的规划和指导，解决在实现微创手术方面已确定的挑战，以及相关研究课题的潜力，包括物理建模及其在计算机辅助干预中的适用性。5)任何涉及的公司或合作者不适用。

1) Brief description of the context of the research including potential impactApplying modern deep learning methodologies in medical imaging applications has shown potentials to automate and to further improve a wide range of clinical applications such as anatomy recognition, localization and pathology quantification. However, requiring large amount of expert labelled training data, uncertainties associated with the labels and difficulties in interpreting the predictions have hindered the clinical adoption of these approaches. This project investigates a meta-learning strategy that can generalise the benefits from weakly-supervised, semi-supervised and unsupervised techniques to imaging applications for prostate cancer and lung disease diagnosis, overcoming these limitations.2) Aims and Objectives -The specific objectives are to:This project aims to develop novel meta-learning methods to address the generalisation issue in deep-learning-based medical image analysis, in particular, incorporating weak supervision, semi-supervision and unsupervised algorithms in meta-learning to minimize the dependency on subjective clinical labelling.The proposed research will focus on learning underlying supervision signals together with intrinsic data structures in both task- and meta-optimization stages. It is the central hypothesis in this project that, by combining the two complementary perspectives of learning strategies, medical image analysis application will benefit the most: while meta-learning utilises its potential to learn "prior" knowledge between tasks, the not-so-supervised algorithms learn the "inner" data knowledge between the "inner-loop data". Not only to overcome the limitation on labelled data, a more interesting direction is to quantify the representation, learned to improve small-data training, to discover the reasoning of the predictions or link these predictions to clinically interpretable measurements. 3) Novelty of Research MethodologyThis project plans to investigate the roles of weak supervision, semi-supervision and unsupervised algorithms in meta-learning, exploring underlying supervision signals in two nested loops. In the inner loop, "pseudo-labels" learned from unsupervised methods will be used to enhance the supervision, improving the performance for target tasks. In the outer loop, task similarities will be investigated using metric learning, improving the generalization to target tasks. Moreover, correlation analysis will be used to quantify the data structure and similarity to known interpretable knowledge, such as clinical measurements and label quality. This helps to measure model and data uncertainties and interpret machine predictions, such as radiologist annotations and scores. 4) Alignment to EPSRC's strategies and research areasThis project aligns with EPSRC's 'healthcare technologies' theme and 'artificial intelligence technologies' research area. Through this project, we aim to address the 'frontiers of physical intervention' challenge that has been laid out as part of the healthcare technologies theme strategy. The project will also address identified challenges in enabling minimally invasive procedures through medical image-based planning and guidance and potential the related research topic includes physical modelling and its applicability in computer aided intervention.5) Any companies or collaborators involvedNot applicable.