The UW’s Machine Learning for Medical Imaging Initiative (ML4MI) selection committee recently announced the winners of the ML4MI Pilot Research Grants. These grants are expected to expand the scope of Machine Learning activity on campus and to encourage cross-disciplinary collaborations to advance the rapidly growing field of machine learning in medical imaging . A central goal of these grants is to spark enduring collaborations that will lead to external funding for further research.
The two winning proposals are “Patient specific hemodynamics using machine learning based fusion of MRI measurements and computational fluid dynamics” from co-PIs Kevin Johnson, PhD, Alejandro Roldán-Alzate, PhD, and Shiva Rudraraju, PhD, and “DeepRad: An accessible, open-source tool for deep learning in medical imaging” with co-PIs Alan McMillan, PhD, and Varun Jog, PhD.
The Pilot Research Grants are specifically targeted for research collaborations between the School of Medicine and Public Health and the College of Engineering at the University of Wisconsin. At least one Primary Investigator or co-PI of the research proposals must come from the Departments of Radiology or Medical Physics and one from the College of Engineering.
The patient specific hemodynamics proposal presented by Drs. Johnson, Roldán-Alzate and Rudraraju seeks to develop vastly superior estimates of in-vivo blood velocities using an artificial intelligence based combination of computer simulations and measurements from an MRI scanner. Clinicians often use measures of blood velocities in the heart and arteries to plan surgical procedures and diagnose disease. MRI provides methods to directly measure 3D velocity fields but produces images that are corrupted by noise and have limited resolution. The interdisciplinary team aims to develop a technique which harnesses modeling, specifically computational fluid dynamics (CFD), to produce images with higher resolution and less noise.
With the DeepRad proposal, Drs. McMillan and Jog want to address the limitations on the accessibility of deep learning techniques for physicians and researchers who do not have expertise or experience in computer programming, machine learning or data science. “While progress has been made in making deep learning tools easier to use, applications still currently require significant software development skills; and although collaborations between medical imaging and computer science researchers are ideal, the rapid maturation of high-performing deep learning tools provided freely by third parties suggests that many applications no longer require significant technical innovation in order to be applied,” said co-PI Varun Jog, PhD. They will leverage deep learning techniques that have been successfully implemented in non-medical data and apply them to medical image data. This includes changes such as extending deep learning frameworks that have been built to use 2-dimensional images, in order to apply them to the 3D and 3D + time datasets that are common in medical imaging. By creating an open-source, fully graphical and user-friendly tool, McMillan and Jog expect to increase the accessibility of deep learning methods that are proving increasingly important in the detection, classification, and diagnosis of disease.
The University of Wisconsin is one of the institutions leading the charge to explore this exciting technology and the impact it promises to have on science and medicine. Team efforts in this area leverage multiple areas of expertise, including clinical imaging, imaging physics, computer science electrical and computer engineering, bioengineering and biostatistics. The UW has outstanding strengths in all these areas, as well as a long and successful history of collegial collaborations between departments.
Author: Chris Temme