Prof. Dr. Florian Knoll

The majority of diagnostic medicalimaging pipelines follow the same principles: raw measurement data is acquiredby scanner hardware, processed by image reconstruction algorithms, and thenevaluated for pathology by human radiology experts. Under this paradigm, every stephas traditionally been optimized to generate images that are visually pleasingand easy to interpret for human experts. However, raw sensor information thatcould maximize patient-specific diagnostic information may get lost in thisprocess. This problem is amplified by recent developments in machine
learning for medical imaging. Machine learning has been used successfully inall steps of the diagnostic imaging pipeline: from the design of dataacquisition to image reconstruction, to computer-aided diagnosis. So far, thesedevelopments have been disjointed from each other. In this project, we willfuse machine learning for image reconstruction and for image-based diseaselocalization, thus providing an end-to-end learnable image reconstruction andjoint pathology detection approach that operates directly on raw measurementdata. Our hypothesis is that this combination can maximize diagnostic accuracywhile providing optimal images for both human experts and diagnostic machinelearning models.

The broad mission of our Center for Advanced Imaging Innovation and Research (CAI2R) is to bring together collaborative translational research teams for the development of high-impact biomedical imaging technologies, with the ultimate goal of changing day-to-day clinical practice. Technology Research and Development (TR&D) Project 1 aims to replace traditional complex and inefficient imaging protocols with simple, comprehensive acquisitions that also yield quantitative parameters sensitive to specific disease processes. In the first funding period of this P41 Center, our project team led the way in establishing rapid, continuous, comprehensive imaging methods, which are now available on a growing number of commercial magnetic resonance imaging (MRI) scanners worldwide. This foundation will allow us, in the proposed research plan for the next period, to enrich our data streams, to advance the extraction of actionable information from those data streams, and to feed the resulting information back into the design of our acquisition software and hardware. Thanks to developments during our first funding period, we are now in a position to question long-established assumptions about scanner design, originating from the classical imaging pipeline of human radiologists interpreting multiple series of qualitative images. We will reimagine the process of MR scanning, leveraging our core expertise in pulse-sequence design, parallel imaging, compressed sensing, model-based image reconstruction and machine learning. We will also extend our methods to complex multifaceted data streams, arising not only from MRI but also from Positron Emission Tomography (PET) and other imaging modalities, as well as from diverse arrays of complementary sensors.

We propose a novel way of reconstructing medical images rooted in deep learning and computer vision that models the process how human radiologists are using years of experience from reading thousands of cases to recognize anatomical structures, pathologies and image artifacts. Our approach is based on the novel idea of a variational network, which embeds a generalized compressed sensing concept within a deep learning framework. We propose to learn a complete reconstruction procedure, including filter kernels and penalty functions to separate between true image content and artifacts, all parameters that normally have to be tuned manually as well as the associated numerical algorithm described by this variational network. The training step is decoupled from the time critical image reconstruction step, which can then be performed in near-real-time without interruption of clinical workflow. Our preliminary patient data from accelerated magnetic resonance imaging (MRI) acquisitions suggest that our learning approach outperforms the state-of-the-art of currently existing image reconstruction methods and is robust with respect to the variations that arise in a daily clinical imaging situation. In our first aim, we will test the hypothesis that learning can be performed such that it is robust against changes in data acquisition. In the second aim, we will answer the question if it is possible to learn a single reconstruction procedure for multiple MR imaging applications. Finally, we will perform a clinical reader study for 300 patients undergoing imaging for internal derangement of the knee. We will compare our proposed approach to a clinical standard reconstruction. Our hypothesis is that our approach will lead to the same clinical diagnosis and patient management decisions when using a 5min exam. The immediate benefit of the project is to bring accelerated imaging to an application with wide public-health impact, thereby improving clinical outcomes and reducing health-care costs. Additionally, the insights gained from the developments in this project will answer the currently most important open questions in the emerging field of machine learning for medical image reconstruction. Finally, given the recent increase of activities in this field, there is a significant demand for a publicly available data repository for raw k-space data that can be used for training and validation. Since all data that will be acquired in this project will be made available to the research community, this project will be a first step to meet this demand.

Public Health Relevance

The overarching goal of the proposal is to develop a novel machine learning-based image reconstruction approach and validate it for accelerated magnetic resonance imaging (MRI). The approach is able to learn the characteristic appearance of clinical imaging datasets, as well as suppression of artifacts that arise during data acquisition. We will test the hypotheses that learning can be performed such that it is robust against changes in data acquisition, answer the question if it is possible to learn a single reconstruction procedure for multiple MR imaging applications, and validate our approach in a clinical reader study for 300 patients undergoing imaging for internal derangement of the knee.