2021 Seed Grants Now Open

Deep learning (DL) models have demonstrated their potential to diagnose disease on medical imaging with performance approaching or even exceeding that of a human radiologist. But deep learning models suffer from a fundamental problem: they often adopt unwanted biases found within the data on which they’re trained. Ultimately, these biases can lead to problems such as discrimination and underserving or underperforming on underrepresented populations.

To make these models more equitable, the research team will quantify the impact of chest x-ray datasets with imbalances in demographics, such as sex and race/ethnicity, on the development of biased DL models for medical imaging diagnosis. Based on these findings,  the team will develop and validate methods to train DL models resistant to demographic imbalances. Ultimately, this work will set the foundation for unbiased DL algorithms which can be safely deployed in diverse patient populations.

Research Team: 

• Paul Yi
• Jeremias Sulam
• Ilya Shpitser
• Cheng Ting Lin

.

The COVID-19 pandemic has exposed longstanding health inequities in United States; evidence is mounting that some ethnic groups are suffering far worse outcomes, including mortality, compared to others. Precise mechanisms driving these disparities are not currently well understood, and may include socioeconomic factors, prevalence of important comorbidities, or even other mechanisms such as differential adherence to government interventions.

In this proposed study, the team will leverage modern causal inference methodology to investigate mechanisms behind observed disparities associated with causal pathways mediated by by socioeconomic factors, biological factors, or differences in care. Their findings will be made available to the scientific and legislative communities as soon as possible to help guide continued research and political intervention to mitigate the causes of inequitable health outcomes in COVID-19.

*Note: this project is a continuation of work started at the center’s fall symposium “Disparities and COVID-19: A Two Day Virtual Workshop on Data Science for Social Good.”

Research Team: 

• Mathias Unberath
• Ilya Shpitser
• Scott Levin
• Jeremiah Hinson

Neovascular age-related macular degeneration (NVAMD) is the leading cause of central vision loss in US adults over 50 years of age. Although optical coherence tomography (OCT) has revolutionized the diagnosis and management of this disease, OCT analysis still has limitations; commonly-used OCT metrics, including central subfield thickness (CST), are not good predictors of functional outcomes like visual acuity, or the clarity and sharpness of one’s vision.

A promising strategy to create a clinical tool that can predict function based on OCT images is to harness the power of deep learning. The research team plans to develop an alternative OCT metric by training a deep learning algorithm that more closely correlates structure (OCT image) with visual acuity. Such a meaningful metric will provide a useful secondary endpoint in clinical trials for new NVAMD therapies, and provide clinicians a new tool for tracking progression/improvement in NVAMD patients.

Research Team: 

• Craig Jones
• T.Y Alvin Liu
• Peter Campochiaro
• Anam Akhlaq

According to the National Brain Tumor Society (1), an estimated 700,000 people in the United States are living with a primary brain tumor and about 80,000 people in the U.S. are diagnosed with a primary brain tumor each year. As a result, large numbers of people must undergo brain surgery every year. Current state-of-art for presurgical, noninvasive planning of tumor removal is to perform blood oxygen level dependent (BOLD) functional magnetic resonance imaging (fMRI).

During scanning, a patient performs a specific task to localize brain functional regions related to the task. A decrease in the relative concentration of deoxyhemoglobin in active cortex reduces the T2/T2* shortening effects of deoxyhemoglobin with resultant net increase in the BOLD signal in activated areas. fMRI necessitates image acquisition methods, which are sensitive to changes in T2* and T2, have sufficient spatial resolution to cover the entire brain, and have sufficient temporal resolution to detect changes in BOLD signal associated with specifics tasks. The spatial resolutions of clinical fMRI presurgical mapping are usually between 8 mm3 and 64 mm3 (i.e. 2–4 mm along each voxel dimension). Given that human cortex is about 3 mm thick, acquiring functional mapping signals with the highest possible fidelity to the underlying neuronal processes is critical for presurgical mapping. With the advent of ultrahigh magnetic field human MRI scanners, i.e., 7T and above, recent studies have probed into depth-dependent cortical mapping with high resolution functional MRI of human brain. The purpose of this study, therefore, is to enhance resolution of 3T fMRI using artificial intelligence on 7T fMRI for the overall goal of depth-dependent cortical mapping in clinical fMRI.

Research Team: 

• Shruti Agarwal
• Haris Sair

More than 300,000 patients in the United States undergo cardiac surgery each year, and substantial morbidity and mortality are common. Acute kidney injury(AKI) is one of the most frequent post-surgical complications, with estimates ranging from 20-50% depending on the procedure. AKI is strongly correlated with reduced quality of life, progression to renal-replacement therapy (i.e. dialysis),and mortality. The pathophysiology of AKI is complex and poorly understood, but likely includes hemodynamic changes (low cardiac output, blood pressure), mechanical factors (emboli, venous congestion), and other mechanisms (vasoconstriction, drugs, inflammation)—many of which are highly modifiable. One of the key challenges to developing preventative strategies for AKI is our inability to monitor intraoperative kidney function.

This overarching goal of this proposal is to develop an automated platform to continuously forecast the risk for AKI based on real-time intraoperative physiological data. This information would allow clinicians to optimize intraoperative variables on-the-fly to reduce the likelihood of kidney injury.

Research Team: 

• Archana Venkataraman

Radiation therapy is a clinically popular treatment for cancer, and knowledge-based radiation therapy treatment planning is becoming an integral part of clinical practice.

However, there is considerable variation in clinical guidance—which identify the feasibility constraints for the plans—of treatment plans among different institutes, hospitals, and oncologists.

In this proposed study, we develop a novel methodology that combines machine learning and inverse optimization to infer utility functions and hidden constraints of a black-box decision-making process. We will evaluate the method in the context of personalized radiation therapy treatment planning, building on a large dataset of historical radiation therapy treatment plans that were clinically approved and delivered. Using a data-driven approach, we apply our iterative approach of machine learning and inverse optimization on the radiation therapy plans to find the underlying criteria that the oncologists considered when approving the plans. Once these criteria are known, a forward optimization model can be used to design personalized treatment plans automatically.

Research Team: 

• Kimia Ghobadi

Older adults are highly susceptible to postoperative complications following major surgery. A growing body of research shows that physical conditioning prior to surgery, often called “prehabilitation,” can provide significant benefit to patients, including reduced complications, shorter recoveries, and lower costs. However, tools to measure the effectiveness of prehabilitation are currently lacking.

Anton Dahbura, MCEH associate research professor, is looking to wearable technology for new insights on the benefits of prehabilitation. His team has developed a platform that can collect data from wearable activity monitors and display feedback to patients on their smartphones. By continuously monitoring patient activity with wearable devices, the team hopes to apply analytics to determine the true relationship between pre-surgery activity levels and surgical outcomes.

The team’s initial study will focus on older adults (over the age of 65) who are preparing for major abdominal colorectal surgery.

Research Team: 

• Anton Dahbura
• Erik Hoyer
• Daniel Young