Hardy, Peter A

Peter Hardy – Radiology, College of Medicine

Research Activities and Computational Needs

The first project we would use the HPC system for is the development of a series of Deep Learning models for medical image segmentation. The specifics of the project are outlined in this excerpt from a recently submitted grant application.

Research objectives & Specific Aims

Our central hypothesis in this project is that we can develop an Artificial Intelligent (AI) algorithm which can segment the major muscles in the thigh with an accuracy equivalent to human operators. This algorithm will enable our current and future research. We will test this hypothesis by accomplishing these specific aims.

Aim 1. We will develop an AI algorithm to segment the vastus lateralis (VL) quadriceps muscle from MR images spanning the entire thigh from the knee to the hip. Accomplishing this aim will create an algorithm we can use to speed up our analysis, and facilitate new analyses, of images acquired in on-going clinical trials in which the thigh is imaged to assess the impact of a traumatic injury.

Aim 2. We will develop an AI algorithm to segment the major muscles in the thigh (the quadriceps: rectus femoris, vastus lateralis, vastus intermedius, vastus medialis, and the biceps and semitendinosus) over the length of the thigh. The quadriceps, which are positioned anteriorly, and the biceps and semitendinosus, which are positioned posteriorly, appear to be affected differently by a traumatic injury.

Significance

The development of this algorithm will allow automated identification of the major muscles in the thigh. This is valuable as it will enable the objective comparison of several measurements of the effects of injury and rehabilitation. The algorithm will enable us to estimate the volumes of the major muscles in the thigh and thus to determine how different muscles atrophy and remodel their extracellular matrix in response to an injury¹. This tool will be central to facilitating a variety of measurements which we would like to propose to support future grant applications to study the effects of traumatic and progressive conditions affecting the musculoskeletal system. 

Individual Projects

Approach. We will develop Deep Learning (DL) algorithms based on manual outlines of muscles. Additional segmentations will be generated by undergraduate  students using the publicly-available and NIH-supported MIPAV software. We have recruited many such students from those aspiring to join the physical therapy program. They are interested in human anatomy and are motivated to work on a research program. The students will be trained and supervised by Meredith Owen, PhD who is a post-doctoral scholar in the department of Rehabilitation Sciences and working under the guidance of Brian Noehren, Professor of Rehabilitation Sciences.  Dr. Owen will train the students to identify the various muscles and how to outline them using the MIPAV software. The DL algorithm will be developed on the high-performance computing system available to all faculty members through the Center for Computational Science (CCS). The goodness of the algorithms will be assessed using traditional measures of segmentation accuracy such as the Dice coefficient and the Hausdorff distance⁴. We will be guided in the development of the algorithm by Dr. Cody Bumgardner who has extensive experience in the development of DL algorithms.

The algorithms will be developed in a step wise approach starting by developing an algorithm to identify the VL on a single mid-thigh image and progressing to identifying the major muscles in the mid-thigh and eventually to segmenting all the major muscles along the entire thigh. This measured approach will allow us to identify and correct problems which might limit the performance of the full DL algorithm. Our goal is to create a DL algorithm which generates the most accurate segmentation possible.

Project Milestones starting from the beginning of the grant funding.

Milestone 1. Within two months we will develop and test a DL algorithm to segment the vastus lateralis (VL) on a single mid-thigh T1 image. The images and the manual segmentations are extant to begin this work.

Milestone 2. Within four months we will develop and test a DL algorithm to segment the quadriceps, the biceps and semitendinosus on a single, mid-thigh, T1 image. The images exist and additional drawing can be accomplished within this time period to accomplish this task.

Milestone 3. Within eight months we will develop and test a DL algorithm to segment the VL over the entirety of the thigh. Additional operators will be recruited to complete the manual segmentation needed to generate the DL algorithm.

Milestone 4. Within 12 months we will test a DL algorithm to identify the quadriceps, the biceps and semitendinosus and four quadriceps muscles over the entirely of the thigh. We will collect all available images and recruit additional operators to generate the required data to form the DL model.

Outcome:

The development of the proposed DL algorithm will enable the evaluation of muscles which affect a human’s gait and performance in sports. The algorithm can be used immediately to enhance our research into the effects of traumatic injuries and therapies to alleviate them as well as future studies such as we are planning to evaluate the development of fibrosis. We are actively developing this grant which would incorporate MR imaging as the central technique and targeting Feb 2023 as the likely date for submission. The Clinical Translational Imaging Science (CTIS) study section and NIAMS institute are the most appropriate for this project. This project will enhance our competitiveness to secure this funding.

Personnel:

Peter A Hardy, PI, PhD, Associate Professor, Department of Radiology, College of Medicine

Students:

There will be undergraduate students working on the project. At this point there are still to be named.

Collaborators:  All collaborators are from the University of Kentucky

Cody Bumgardner, UKY, PhD Assistant Professor, Department of Pathology, College of Medicine

Brian Noehren, PhD, Professor, UKY, Department of Rehabilitation, College of Health Sciences

Christopher Fry, PhD, Associate Professor, UKY, Department of Athletic Training, College of Health Sciences

Meredith Owen, PhD, Post Doctoral Fellow, UKY, Department of Rehabilitation, College of Health Sciences

Computational methods

The overall goal is to develop a convolutional neural network which is able to provide accurate segmentation of magnetic resonance images of the thigh in order to identify specific muscles in the thigh.