Innovations in Biomedical Imaging Technology
Researchers in the Department of Radiology at NYU Langone advance biomedical imaging through technological innovation. The imaging devices and methods widely used in healthcare today are sophisticated products of interdisciplinary scientific and engineering gains in physics, mathematics, materials science, electrical engineering, computer science, biochemistry, and medicine.
We develop new technologies in several areas to further enhance the value of medical imaging by making it more informative for doctors, more comfortable for patients, simpler for technicians, and faster for everyone.
Artificial Intelligence in Imaging
We are at the forefront of developing artificial intelligence (IA) in biomedical imaging and conduct AI research that spans the spectrum of radiology practice and addresses every step of the imaging process, from data acquisition to image formation to image interpretation.
Our research has the potential to dramatically increase the speed of MRI and to significantly expand access to this valuable imaging modality. We train neural networks to guide radiofrequency pulse sequences, reconstruct images from raw MRI data, and search for signatures of disease in the patterns of MR signal itself.
Our investigators use machine learning to design and train neural networks in image classification in order to discover medically relevant information, some of which may not be visible to the naked eye. We investigate AI methods to increase diagnostic consistency and assist clinical radiologists in reading and evaluating exams. Specifically, our research is aimed at more accurate diagnosis of breast cancer, better classification of neurological conditions, and early detection of diseases that have slow, asymptomatic onset, such as osteoarthritis and Alzheimer’s disease.
Apart from exploring the potential of machine learning to improve the acquisition, formation, and classification of medical images, we are also engaged in evaluating how algorithms may enhance radiology operations and clinical workflow.
Intelligent Image Acquisition, Reconstruction, and Analysis
Our researchers are developing new technologies that move biomedical imaging toward fast, simple, universal acquisitions that yield quantitative parameters sensitive to specific disease processes.
Conventional imaging too often relies on start-and-stop scanning protocols, long scan times, and complex scan settings aimed at obtaining specific types of contrast. We envision a different paradigm in which rapid, continuous, comprehensive scans are automatically tailored to each patient and deliver rich information about properties of tissue, biophysical processes, and signatures of disease.
Our core expertise in pulse-sequence design, parallel imaging, compressed sensing, model-based image reconstruction, and machine learning, together with the most recent advances in the field—including ones made by our research group—allows us to question basic assumptions about scanner design and the classical imaging pipeline that produces series of qualitative images for human radiologists to interpret.
Some of our continuous imaging techniques are already available on MRI systems sold around the world. As we expand and improve these methods, we are also working on intelligent pulse sequences that optimize image acquisition in real time and on deep learning image reconstruction, which we believe can accelerate MRI scans 10-fold.
We are also investigating ways of enriching imaging data with quantitative information to assist diagnosis and therapy and to further guide intelligent acquisition, reconstruction, and analysis methods.
Flexible MRI Technologies
We are developing flexible imaging hardware, creating methods to navigate rather than control complex scanner environments, and making MRI systems more intelligent and informative.
Our research is aimed at relaxing constraints on medical imaging hardware while delivering better information. For example, ultraflexible high-impedance MRI coils invented by our team have overturned longstanding design limits on imaging arrays. In contrast to conventional arrays, which are set in rigid geometric arrangements, high-impedance coils can flex with the anatomy, making possible MRI studies of biodynamics of moving joints.
We also proposed a new method for multiparametric ultra-high-field MRI called plug-and-play MR fingerprinting (PnP-MRF). This method takes advantage of principles derived from MR fingerprinting to leverage magnetic field inhomogeneities rather than taking pains to control them, which has been the conventional approach. Our method is now in a translational research phase and is available as a work-in-progress package on select Siemens scanners.
Our scientists explore how advances in sensor miniaturization and AI may augment low-field MRI and perhaps even refashion scanner design. To that end, we operate a test bay with a research-only Siemens MRI system ramped down to 0.55 Tesla (0.55T) and dedicated to investigations into multisensory imaging.
We also study potential imaging uses of high-permittivity materials for ultra-high-field MRI, construct biomorphic phantoms that mimic tissue properties, engineer highly specialized multinuclear coils for metabolic imaging, and build state-of-the-art clinical coils for routine medical applications.
Microstructural Information from MRI Scanners
We are developing imaging techniques to uncover information about tissues at the cellular level, also called the mesoscale, where disease processes often originate but evade detection.
Although tissue microstructure lies beyond the image resolution limit of MRI, properties of the cellular environment do inform MR signal, which originates at an even smaller scale of molecules. We develop sophisticated biophysical models, imaging methods, and validation strategies to identify signatures of specific microstructural features in MR signal and to create clinically viable ways of mapping such properties.
Our researchers have published pioneering work that elucidates the mesoscale origins of diffusion-weighted MRI data and identifies parameters within which the data carry reliable information. In deriving new functional forms that capture such information, we use methodology borrowed from fundamental physics. To validate theoretical developments, we employ leading-edge validation tools, including electron microscopy and computational neuroanatomy. Our scientists have also developed state-of-the-art denoising methods.
In particular, we develop and validate tissue- and disease-specific models of neurodegeneration, muscular disorders, and cancer. These models are intended to provide the physical and biological groundwork for the development of imaging methods across other areas of our research. We are also exploring several test cases in which microstructure mapping may be used to inform diagnosis and therapy, such as to reduce overtreatment of prostate cancer or to improve neurosurgical planning.
Multifaceted Imaging Approaches
We develop information sources complementary to conventional MRI in order to enrich imaging data and expand the sensing strategies relied on in acquisition of clinically relevant knowledge.
Our scientists are developing joint image reconstruction methods for simultaneously acquired PET/MR data in order to take advantage of synergies between the two modalities. We have developed novel devices and techniques for correcting the effect of respiratory motion and for precisely co-registering PET and MR data. Several of these technologies are now in the translational research phase, with prototypes and work-in-progress packages available on select Siemens systems. We are also developing radiotracers that specifically leverage simultaneous acquisition and joint reconstruction, and that enable validation of integrated physiological monitoring.
Our researchers create new and investigate existing PET radiotracers to better understand molecular mechanisms of pathology as well as the mechanisms of drug action in various types of cancer and neurological disease.
We also develop advanced MRI methods, such as multinuclear imaging, MR spectroscopy, and MR fingerprinting, to investigate biometabolism, support diagnosis, and inform therapy for conditions like traumatic brain injury, stroke, cancer, osteoarthritis, and diabetic neuropathy. For example, our scientists are using multiparametric MR spectroscopy to detect early-stage neurodegeneration in normal-appearing white matter and the thalamus of recently diagnosed patients with relapsing–remitting multiple sclerosis. In preclinical Alzheimer’s disease, we are investigating whether proton MR spectroscopy can improve the predictive value of current imaging biomarkers, since changes in metabolism are usually the earliest manifestations of disease.
To carry out these investigations, we often design, engineer, and build dedicated state-of-the-art multinuclear coils tuned to the frequencies of two or more elements of interest. Dual-, triple-, and quadruple-tuned coils allow us to gather data about several metabolites during a single scan. We also develop custom post-processing techniques for data analysis.
Recent advances in computing power, machine learning, and sensor miniaturization have created an opportunity to augment traditional MRI acquisitions with new, informative data. Our scientists are exploring the possibilities of outfitting scanners with non-traditional technologies to create multisensory imaging machines. In one such approach, we developed a portable stand-alone sensor to detect patient respiratory motion and include the resulting data in the MR acquisition. The approach, called Pilot Tone, can be used to correct images for the effects of breathing motion and is now being tested on select Siemens systems.