Segmentation of Clinically Acquired Medical Images
Brief Description of Project:
Project Area 1:
Ventral hernia (VH) repair is one of the procedures most commonly performed by general surgeons worldwide, yet extensive variation exists in its delivery. Optimal candidates for repair, repair type, and ideal use of hernia mesh remain to be defined. There currently does not exist a widely accepted VH classification system to help identify disease severity and to improve outcomes.
Even under optimal conditions, VHs occur in up to 28% of over 2 million patients undergoing abdominal operations each year. With current best-practices, VH repair is fraught with failure – recurrence rates ranging from 24-43% are reported. Recurrence of previously repaired VHs increases morbidity to patients and increases healthcare costs. Meanwhile, VHs continue to rise in incidence with nearly 350,000 repairs performed and total procedural costs for VH repair of $3.2 billion in the United States in 2006.
Most VH patients undergo computed tomography (CT) scanning to evaluate their abdominal wall. This information is used to make clinical judgments about a particular patient's hernia for treatment and prognosis. Currently, these decisions are not evidence-based, but rather subjective and based mostly on expert opinion. We hypothesize that the CT datasets obtained from these studies are underutilized and provide a potentially rich — and automated — means of better characterizing VH.
We propose to translate recent advances in abdominal image processing and imaging informatics to link a CT-based VH classification with treatment options and outcomes. This effort is an equal collaboration between surgical and engineering sciences by two early stage new investigators. Specifically, we will: (1) Create a library of expertly labeled abdominal wall structures and VH defects on CT. (2) Develop and evaluate algorithms for the automated labeling of abdominal wall structures, VHs, and bony anatomy on clinically acquired CT data for planning of VH repair. (3) Assess the prognostic value of image-derived VH metrics for intra-operative, 30-day, and one-year outcomes. Metrics will include the area of disruption, volume of herniated tissue, degree of impact to muscles (ruptured versus displaced), and displacement/volumetric changes in abutting wall structures versus relatively normal tissues. The imaging data, which are already routinely collected, will help identify patients who may benefit from specific interventions and ultimately reduce the chance of hernia recurrence and morbidity. These advances will enable exploration of biomarker development and tissue modeling to improve the precision and quality of VH patient care.
Project Area 2:
$100 billion in direct and indirect economic costs. About 75% of civilian TBIs that occur each year are classified as mild TBI (mTBI), which is also one of the most common consequences of military deployment. Over three-quarters of mTBIs show no visual abnormalities on modern imaging sequences, but these injuries have real cognitive, emotional, and psychological impacts.
Computed tomography (CT) is the primary modality to stratify TBI severity and predict clinical outcomes, while Magnetic resonance imaging (MRI) is not standard of care in acute mTBI evaluations, despite its potential promise in identifying the subtle physical changes that underlie these mild TBI (mTBI) symptoms. Complicating the subtle mTBI diagnosis is that the human brain undergoes substantial changes during maturation and aging, so injury related brain changes are confounded with natural process during these time periods. It is becoming clear that such "precision medicine" can be realized through improved understanding of large-scale population studies which have the statistical power to reveal nuanced effects. Large scale imaging studies (so called "mega-analyses") have the potential statistical power to model subtle effects and advance imaging-based personalized medicine, but these tools have not been presented.
There are substantial theoretical and technical challenges in performing analysis with an order of magnitude more subjects. Simply representing a data volume of (pixels), let alone performing processing or structured inference, can require special hardware. Although genomic analyses with 104 subjects and 106 markers are not uncommon, we are not aware of any large-scale statistical frameworks that account for the special lattice and spatial dependence properties of imaging data. The primary hypothesis of this work is that statistical inference approaches can be generalized to enable very large-scale study of the heterogeneous imaging archives present at Vanderbilt.
At Vanderbilt, we have the VUIIS Central Experimental Repository Archive and Library (VISCERAL) which contains structural and functional MRI data from nearly 10,000 scans from multiple research studies over the past decade. Recently, this vast archive has been brought under a single IRB research archive to promote collaboration, innovation and discovery. However, this resource has not yet been successfully mined to further clinical or basic science research. In this pilot study, we will create a statistical atlas of multi- modal brain anatomy from more than 10,000 brain scans. We will implement a statistical clustering approach using functional, structural, and quantitative MRI data to reveal multivariate patterns in this data. These efforts will demonstrate our core competencies to address the challenges of large-scale inference and to inform future research projects. We will leverage the results and experience gained from this pilot study in a follow-up NIH/NINDS R01 proposal to use the BioVU and Synthetic Derivative databases to better understand brain injury within the background development and aging.
- Experience with imaging
Nature of Supervision:
Summer internship (full time)
A Brief Research Plan (period is for 10 weeks):
2 weeks - project overview and plan development
6 weeks - project implementation
1 week - statistical evaluation
1 week - project write-up
Number of Slots: 4
Electrical Engineering & Computer Science