TrustMRI: Trustworthy and Robust Magnetic Resonance Image Reconstruction with Uncertainty Modelling and Deep Learning

磁共振成像(MRI)是各种检查的主要诊断手段，因为缺乏电离辐射，而且它能够探测生理的各个方面。近年来，MRI在英国的使用有了很大的增长，随着技术的进步和人口的老龄化，这一需求可能会继续同比增长。然而，遗憾的是，MRI数据采集过程的物理特性使其固有的速度变慢，而且对MRI的需求持续增加，可靠性降低，也导致患者等待时间更长，程序重复。因此，社会必须找到新的方法来改进和优化高效的磁共振成像工作流程。最近，人工智能(AI)技术为大大加快MRI采集过程提供了可能，并使传统重建方法的局限性得到了突破。然而，对于它们的可信性和对未见病例的失败管理仍然缺乏考虑，这限制了它们在临床实践中的翻译潜力。随着基于深度学习的MRI重建技术的日益发展，对深度学习重建扫描的可信性和不确定性的认识变得越来越必要，并且对下游诊断决策也是至关重要的。本项目旨在解决AI支持的MRI重建中日益严重的AI可信性问题。这项拟议的研究将整合和推进机器学习和医学成像领域的最先进研究。它将开发新的贝叶斯深度学习方法来量化模型驱动的MRI重建的不确定性，建立原始的故障预测机制来评估不确定性，并研究先进的测试时间不确定性减少技术来处理不分布的数据。这将有助于创建一个简化的管道，以促进基于深度学习的磁共振成像重建中常见的不确定性做法。它还将在两个临床应用上进行评估，即加速病理性脑MRI和运动破坏的心脏MRI重建。人工智能MRI重建的发展及其翻译需求的融合打开了令人兴奋的可能性，我们建议在这个项目中进行研究。

Magnetic resonance image (MRI) is the leading diagnostic modality for a wide range of exams due to the lack of ionising radiation and its ability to probe various aspects of the physiology. The use of MRI in UK has seen a large increase in recent years and with the technological advances and an ageing population, this demand is likely to continue to increase year-on-year. However unfortunately the physics of MRI data acquisition process makes it inherently slow, and the sustained increase in demand for MRI and its reduced reliability have also led to patients' longer waits and repeated procedures. It is therefore essential that society finds new ways to improve and optimise towards efficient MR imaging workflows. Recently, artificial intelligence (AI) techniques have opened the possibility to accelerate the MRI acquisition process considerably and have enabled progress beyond the limitations of conventional reconstruction methods. However, there is still a lack of consideration of their trustworthiness and failure management on unseen cases, which limits their translational potential in clinical practice. With the increasing development of deep learning-based techniques for MRI reconstruction, awareness about trustworthiness and uncertainty over deep learning reconstructed scans are becoming necessary and are also critical for downstream diagnostic decision-makings.This project aims to tackle the critical and growing problem of AI trustworthiness for AI-enabled MRI reconstruction. The proposed research will integrate and advance state-of-the-art research in machine learning and medical imaging. It will develop novel Bayesian deep learning approaches to quantify uncertainty for model-driven MRI reconstruction, build original failure prediction mechanisms to evaluate uncertainty, and investigate advanced test-time uncertainty reduction techniques for handling out-of-distribution data. This will conduce to creation of a streamlined pipeline to foster the common uncertainty practices in deep learning-based MRI reconstruction. It will also be evaluated on two clinical applications of accelerated pathological brain MRI and motion-corrupted cardiac MRI reconstruction. The confluence of the development in AI-enabled MRI reconstruction and its translational need opens exciting possibilities that we propose to investigate in this project.