The AIMI Seed Grant program seeks to stimulate and support the creation of innovative and high-impact ideas that will advance the fields of medicine & imaging. Of 45 applications received for the 2019 call for proposals, AIMI is funding these seven exciting and innovative projects based on their scientific merit, potential clinical impact, and alignment with program priorities.
Artery Occlusion Detection on Digital Subtraction Angiography in Acute Ischemic Stroke Patients
PI: Greg Zaharchuk
Co-Investigators: Yannan Yu, Huy Do, Jiahong Ouyang
The specific aims for this project are to build a database with digital subtraction angiography (DSA) images and expert annotations for acute ischemic stroke cases; train a neural network to detect large vessel occlusions on DSA; and fine-tune the trained neural network to detect medium and small vessel occlusions. Ultimately, this tool could be integrated into the angiography units to provide real-time guidance for thrombectomy cases, and assist the decision-making and improve the quality of care for acute ischemic stroke patients.
Automating the Analysis of Echocardiographic Images for Predicting Post-Operative Right Heart Failure
PI: William Hiesinger, Jeffrey Teuteberg
Predicting post-operative right ventricular failure (RVF) has proven challenging despite decades of research. We aim to use a neural network to segment transthoracic echocardiograms (TTEs) collected as part of standard clinical care. The segmented TTEs will be used for the extraction of novel RVF metrics that have shown promise in predicting post-operative RVF, which will then be used to generate predictive models that are both sensitive and specific for post-operative RVF. The outcome of this research project will be the creation of a validated, clinically translatable, end-to-end automated tool for predicting postoperative RVF in patients undergoing LVAD implant.
Computer Vision Prediction of Cardiovascular Outcomes through Echocardiography
PI: Euan Ashley, James Zou
Co-Investigators: David Ouyang, Robert Harrington
Our central hypothesis is that deep learning can identify additional phenotypes of subclinical disease and cardiovascular risk factors not readily detected by human interpreters, and this phenotyping can risk-stratify patients and alter management. The objective of this application is to combine data science, machine learning and cardiovascular imaging to assess how convolutional neural networks can improve the predictive value and accuracy of echocardiography in diagnosing cardiovascular disease.
Data and Supervision: How Much is Enough to Detect a Lethal Killer?
PI: Rusty Hofmann, Christopher Ré
Co-Investigators: Anna-Margaretha Karmann, Jared Dunnmon, Andre Souffrant
This project aims to characterize the number of patients and image slices, and the granularity of supervision needed to generate an accurate ML model to detect life-threatening deep vein thrombosis (DVT) on abdominopelvic and lower extremity CT scans. We will further assess the ability of efficiently-trained, weakly supervised machine learning models to perform on par with their fully supervised equivalents. If successful, this project would result in a clinically valuable system for detecting DVT, highlighting cases with emergent findings for immediate review.
Deep Learning for Safe Watchful Waiting of Low Risk Basal Cell Carcinoma
PI: Eleni Linos, Olivier Gevaert
Co-Investigators: Roberto Novoa, Justin Ko
The goal of this study is to advance the field of medical imaging by adapting our established deep learning algorithm designed for skin lesion diagnosis to categorize and track changes in BCC images over time taken in the home environment. We will then use this adapted algorithm as the basis for a mobile application for capturing home-based imaging and patient-reported data on low-risk BCCs. Long-term, this mobile tool will change the paradigm for treatment of low-risk skin cancer in the elderly by providing the clinical basis for home-based active surveillance of low-risk BCC.
Exploring AI for Blood Testing Using Magnetic Levitational Imaging Cytometry
PI: Utkan Demirci, Gozde Durmus
Co-Investigators: Pranav Rajpurkar
This project aims to explore a new portable imaging system; i.e., levitational imaging cytometry, utilizing a new paradigm of measuring density distribution of cells with an unprecedented precision and simplicity. This will create a new realm in cellular imaging and diagnostics, when integrated with deep learning. Overall goal of this project is to use deep learning to analyze blood samples levitated and imaged in the portable magnetic levitation system developed by our group to achieve a point of care clinical blood lab tool.
Using Machine Learning-based Radiomics to Distinguish Lung Cancer on CT from a Multi-Center VA Cohort
PI: Rajesh Shah, Sandy Napel
Co-Investigator: Olivier Gevaert
While Machine Learning-based Radiomics (MLR) has previously shown good results in differentiating lung cancer from benign pulmonary nodules (BPN) and identifying histologic subtypes, application is limited because model creation generally used data from small localized cohorts. Given the high rate of lung cancer in the VA population, MLR using the VA’s new research centralized imaging database could create a widely applicable algorithm for distinguishing small cell lung cancer (SCLC) from non-SCLC (NSCLC) and BPN.