Reliable AI for Medical Image Reconstruction

项目摘要/摘要 深度神经网络在从计算机到神经网络的各种学科中取得了广泛的经验成功。 视觉到自然语言处理。然而，在诸如磁共振成像（MRI）的医学成像中， 各种挑战，包括缺乏高质量的培训数据，对腐败/异常值缺乏鲁棒性， 训练和测试时间之间的分布变化以及记录的可靠性和可信度的缺乏阻碍了 人工智能的广泛使用和适应。该项目开发新的基于深度学习的架构，算法和 培训机制，处理这些挑战，创造一个新的工具包，磁共振成像重建， 强大，可靠，值得信赖，但可以在多个医院系统中进行协作和私人培训。 使用这个新的工具包，该项目解决了三个主要的MRI重建挑战（1）减少采集 通过更高的加速因子，（2）即使在低强度磁场下， ELD，以及（3）处理运动伪影。与一些国家的肌肉骨骼（MSK）部门合作， 该项目还涉及收集，策划和发布新的数据集和开源 重建软件旨在解决这些关键挑战。这也将有助于吸引进一步的研究， 机器学习/人工智能社区，以进一步改善这一重要的医学成像模式。 医疗保健影响：该项目将显著增强MRI这一重要诊断工具。（一） 采集时间的减少将同时提高诊断的准确性和患者的舒适度。 (2)由低场扫描仪引起的噪声/非线性伪影的减少将导致尺寸的减小， MR扫描仪的重量。这可能最终允许MRI用于护理点或紧急情况 在床边，也开辟了大量的新用例。(3)运动伪影的减少增加了 对各种新疾病和病症的诊断准确性，从而实现MRI的新诊断用例。还有， (1)允许更多的患者使用同一台机器接受扫描，以及（2）降低磁体的成本， MR扫描仪的操作空间。这可以显着降低患者成本，从而增加获得这种 诊断上重要的医学成像模式。此外，对MSK数据收集的关注可以大大提高 有助于准确诊断病理学，如膝关节、盂唇和旋转体中的细微关节撕裂或撕裂 肩关节撕裂，手腕韧带撕裂。对臂丛神经病变的特别关注也是 由于臂丛神经是一种复杂的解剖结构， 为上肢、肩部和上胸部提供神经支配的关键功能。臂丛 MRI研究将使这种复杂的解剖结构的细节与低/传统的电场强度 这在患者治疗中，特别是在创伤环境中是至关重要的。

Project Abstract/Summary Deep neural networks have enjoyed wide empirical success in a variety of disciplines ranging from computer vision to natural language processing. However, in medical imaging such as Magnetic Resonance Imaging (MRI), a variety of challenges including lack of high-quality training data, lack of robustness to corruption/outliers and distribution shifts between train and test time as well as documented lack of reliability and trustworthiness impede the wide use and adaptation of AI. This project develops new deep learning-based architectures, algorithms and training mechanisms that that deals with these challenges creating a new toolkit for MRI reconstruction that is robust, reliable, trustworthy yet can be trained collaboratively and privately across multiple hospital systems. Using this new toolkit the project addresses three major MRI reconstruction challenges (1) reducing acquisition time via higher acceleration factors, (2) enabling high quality reconstruction even with low-intensity magnetic elds, and (3) dealing with motion artifacts. In collaboration with Musculoskeletal (MSK) sections of a few major universities, this project also involves gathering, curating and releasing new datasets and open source reconstruction software aimed at addressing these key challenges. This will also help attract further research from the machine learning/AI community to further improve this important medical imaging modality. Healthcare Impact: This project will signi cantly enhance MRI which is an important diagnostic tool. (1) reductions in the acquisition time will simultaneously increases both the accuracy of diagnosis and patient comfort. (2) the reduction of noise/nonlinear artifacts caused by low- eld scanners will lead to a reduction in the size and weight of the MR scanner. This may eventually allow MRI to be used at point of care or for emergency scenarios at the bedside and also open up a plethora of new use cases. (3) reductions in the motion artifacts increases the accuracy of diagnosis for a variety of new diseases and conditions enabling new diagnostic use cases for MRI. Also, (1) allows more patients to receive a scan using the same machine and (2) lowers the cost of the magnet and the space of operation of MR scanners. This can signi cantly reduce patient cost and thus increase the access to this diagnostically important medical imaging modality. Furthermore, the focus on MSK data collection can greatly facilitate accurate diagnosis of pathology such as subtle meniscal tears or chondrites in the knee, labral and rotator cu tears in the shoulder, and ligament out tears in the wrist. The particular focus on brachial plexopathy is also expected to have signi cant healthcare bene ts as the briachial plexus is an intricate anatomic structure with the critical function of providing innervation to the upper extremity, shoulder, and upper chest. The brachial plexus MRI study will enable great detail of this intricate anatomical structure with low/conventional eld strengths which is of paramount importance in patient treatment, especially in traumatic settings.