Causal Generative Modelling for Identifying and Alleviating Biases of Neural Networks in Medical Imaging

随着深度神经网络(DNNS)的出现，医学图像分析正在经历一场变革，这是一种机器学习模型，越来越多地作为工具被试验到临床工作流中，例如用于辅助病理检测。然而，由于患者群体、采集扫描仪和协议的多样性，医学图像的多样性对确保模型部署后的可靠性能提出了挑战。由于医学影像数据库的大小有限，它们通常不包括这种异质性。因此，训练数据可能包括偏差，例如使用特定类型的扫描仪或主要从特定年龄的患者获取的图像。这些偏差可以由针对该数据训练的模型继承，如果部署该模型来处理具有不同特征的数据，则该模型可能表现不佳。这可能会在医疗保健领域产生严重后果。当前一代的DNN缺乏可靠的机制来检测和提醒用户预测可能不可靠。本项目的目的是为医学成像创建更安全、更可靠的机器学习。我们的方法是开发因果生成模型(CGM)，能够明确地捕捉影响或不应该影响模型预测的因素之间的关系。通过这样做，我们可以准确地找出导致偏差和次优模型预测的特定数据特征。为此目的，该项目有以下具体目标：a)第一个目标是开发CGMS，用于对数据库的特征进行建模并识别其中的任何偏差。医学图像数据集通常包含年龄、性别、疾病严重程度和使用的成像设备等变量之间的相关性，其中某些组合比其他组合出现得更频繁。我们将开发新的基于因果关系的方法来模拟这些变量之间的关系，它们如何影响图像外观，并确定这种关系是理想的还是数据中虚假相关性的伪像。然后，因果推理将能够合成人造图像，即所谓的反事实，用于原始数据库中未被充分代表或缺失的特征。尽管在遗传学和流行病学等其他健康研究领域中很常见，但因果推断在成像方面的使用要少得多。b)第二个目标是开发方法，以确定已经为感兴趣的任务(例如，检测病理)而训练的模型所继承的偏差，这可能是由于训练数据库存在偏见。这对于确定它在部署时的表现是至关重要的，例如，确保在不同的患者人口统计或图像质量下进行公平可靠的操作。c)最终目标是开发技术，以减轻在预先培训的模型中发现的任何偏差。这是对前面目标的补充，并将它们的输出结合在一个框架中：在确定了模型(B)的偏差之后，我们将使用数据的因果生成模型(A)来生成模型往往表现不佳的特征的合成但现实的数据。这些合成数据将用于调整模型，增强其性能，使其更可靠地运行。该项目属于EPSRC医学成像研究领域和医疗保健技术主题。拟议的研究将产生基于因果推理的新技术，以减轻医学图像分析中的偏差，这将导致公平和更可信的模型，这是在真实世界的临床环境中采用此类工具所必需的。

Medical image analysis is undergoing a transformation with the emergence of Deep Neural Networks (DNNs), a type of machine learning models that are increasingly being trialled into clinical workflows as tools, for example to assist pathology detection. The diversity of medical images, however, resulting from the variety of patient populations, acquisition scanners and protocols, poses a challenge for ensuring reliable performance after model deployment. Because medical imaging databases are of limited size, they commonly do not cover this heterogeneity. As a result, training data can include biases, for example images acquired using a specific type of scanner or primarily from patients of specific age. These biases can be inherited by a model trained on this data, which may underperform if deployed to process data with different characteristics. This can have serious consequences in healthcare. The current generation of DNNs lacks dependable mechanisms for detecting and alerting users when predictions may be unreliable.Aim of this project is to create safer and more dependable machine learning for medical imaging. Our approach is to develop Causal Generative Models (CGMs) capable of explicitly capturing relationships between factors that either influence or should not influence model predictions. By doing so, we can pinpoint the specific data characteristics responsible for biases and sub-optimal model predictions. This knowledge can then be used to alert a user of possible sub-optimal model operation, or use this knowledge to alleviate the influence of such data characteristics, to obtain fairer and more reliable models.To this end, the project has the following specific objectives:a) The first objective is to develop CGMs for modelling the characteristics of a database and identifying any biases therein. Medical image datasets typically contain correlations between variables such as age, sex, disease severity and used imaging equipment among others, with certain combinations occurring more often than others. We will develop novel causality-based methods for modelling the relations between these variables, how they affect image appearance, and identify if such relations are desirable or artifacts of spurious correlations in the data. Causal inference will then enable synthesis of artificial images, so called counter-factuals, for characteristics that are under-represented or missing from the original database. Although common in other areas of health research such as genetics and epidemiology, causal inference has seen far less use in the context of imaging.b) The second objective is to develop methods for identifying biases inherited by a model that has already been trained for a task of interest (e.g. detection of a pathology), potentially due to a biased training database. This is critical in order to determine how well it will perform upon deployment, for example to ensure fair and reliable operation under different patient demographics or image quality.c) The final objective is to develop techniques for alleviating any biases identified in a pre-trained model. This complements the previous objectives and combines their output in a single framework: Having identified biases of a model (b), we will use a Causal Generative Model of the data (a) to generate synthetic but realistic data of the characteristics for which the model tends to underperform. This synthetic data will be used for adapting the model, enhancing its performance for more reliable operation. This project falls within the EPSRC Medical Imaging research area and the Healthcare Technologies theme.The proposed research will generate novel techniques based on causal inference for the mitigation of bias in medical image analysis, which will lead to fair and more trustworthy models, necessary for adoption of such tools in real-world clinical settings.