Identifying and alleviating sensitivity of neural networks to distribution shift in medical imaging

机器学习的进展，特别是有效的深度神经网络的发展，有可能改变医学图像分析，从而改善病理学的诊断和治疗。基于机器学习的工具的巨大价值在于加速和提高临床工作流程中所需的常见和繁琐任务的有效性，例如通过促进检测或描绘患者扫描中的异常。然而，在确保这些算法的可靠性方面的挑战阻碍了神经网络在临床上的广泛应用。具体地说，当前用于开发和训练机器学习算法的方法通常假定用于训练模型的数据将具有与将应用该模型的数据相同的特征。如果这一假设不成立，这种现象在技术上被称为“数据分布转移”，那么模型的性能可能会下降。由于医学影像工作流程的性质，这种“分布转移”是不可避免的。例如，可以使用来自不同制造商的扫描仪、使用不同的采集协议或来自不同的人群组来采集图像。由于与医疗保健应用相关的高风险，每一个错误的预测都可能对单个患者造成灾难性的影响。因此，为了使机器学习发挥其潜力，有必要开发识别哪些数据特征可能导致模型性能下降的方法，以及降低相关模型敏感度的技术，以确保基于机器学习的工具在临床工作流程中的可靠和安全部署。本研究将调查成像数据的哪些特征导致ML模型的预测性能在与训练过程中使用的数据特征不同时降级，并开发降低此类敏感度的方法。该项目的一个技术目标是开发新的生成性网络，该网络可以学习训练和测试数据中的特征分布。然后，生成性网络能够生成具有用户指定的特征的新的合成数据。此属性可用于解释和识别哪些类型的成像特征会降低另一感兴趣模型的性能。此外，我们将使用这种生成性模型来减轻对这些特征的敏感性。例如，生成模型可用于创建具有与从部署环境获取的数据的特征相似的特征的合成图像。然后，这些可以用于训练更健壮的模型，从而提高在训练数据之外进行预测的能力。此外，这项研究的目的是开发一种方法，使神经网络能够在部署后自动调整其内部表示，以便对其处理的特定数据而不是对其进行训练的数据更加优化。这将允许更优化的性能，而无需重新训练模型或重新收集与工具要部署到的环境中的数据匹配的数据。最终，该项目将开发下一代框架和技术，以开发更透明和可靠的基于机器学习的医疗保健工具。该项目将先进的计算机科学技术应用于医学应用。它试图使用从医疗数据中提取有用信息的算法来优化疾病诊断、决策和治疗。因此，该项目属于EPSRC医学成像研究领域和医疗保健技术主题。

Advances in machine learning, particularly the development of effective deep neural networks, have the potential to transform medical image analysis and consequently improve diagnosis and treatment of pathology. The great value of machine-learning based tools is in accelerating and improving effectiveness of common and tedious tasks required within clinical workflows, such as by facilitating detection or delineation of an abnormality in a patient's scan. However, wide deployment of neural networks for clinical application is hindered by challenges in ensuring reliable of these algorithms. Specifically, current methodologies for developing and training machine learning algorithms commonly assume that the data used for training the model will have the same characteristics as the data on which the model will be applied. If this assumption does not hold, a phenomenon known technically as "data distribution shift", then model performance is likely to degrade. Such "distribution shift" is unavoidable due to the nature of medical imaging workflows. For example, images could be acquired using scanners from different manufacturers, using different acquisition protocols or from different populations groups. Due to the high risk associated with healthcare applications, every false prediction can be disastrous for an individual patient. Therefore, in order to enable machine learning deliver its potential, it is necessary to develop methodologies for identifying what data characteristics may lead to degradation of model's performance, and techniques for alleviating associated model sensitivity, to ensure reliable and safe deployment of machine learning based tools in clinical workflows.This research will investigate what characteristics of the imaging data cause the predictive performance of ML models to degrade when they differ from the characteristics of data used during training, and develop methods to alleviate such sensitivity. One technical objective of this project is to develop novel generative networks that can learn the distribution of characteristics in the training and testing data. Generative networks are then capable of generating new, synthetic data, with the characteristics that the user specifies. This property can be used to interpret and identify what types of imaging characteristics degrade performance of another model of interest. Moreover, we will use such generative models for alleviating sensitivity to such characteristics. For example, generative models can be used to create synthetic images with characteristics similar to those of data that are acquired from the environment of deployment. These can then be used for training more robust models, with increased capability to make predictions outside the training data. Moreover, this research aims to develop methods that can enable a neural network to autonomously adapt its internal representation after deployment, to be more optimal for the specific data that it processes, rather than those it has been trained on. This would allow more optimal performance without the expense of retraining the model or re-collecting data that match those from the environment that a tool is aimed to be deployed in. Ultimately, the project will develop the next generation of frameworks and techniques for developing more transparent and reliable machine-learning based tools for healthcare. This project applies advanced computer science technologies to medical applications. It tries to optimise disease diagnosis, decision making and treatment using algorithms that extract useful information from medical data. Therefore this project falls within the EPSRC Medical Imaging research area and the Healthcare Technologies theme.