The Defense Advanced Research Projects Agency (DARPA) has selected a Drexel University research team to improve battlefield injury diagnostics by developing point-of-care ultrasound (POCUS) artificial intelligence (AI) innovations. Drexel University is one of five institutions chosen to tackle DARPA’s ground-breaking project.
The 18-month Point-of-Care Ultrasound Automated Interpretation program tasks nationally-reputable researchers with creating adaptive AI models that decipher ultrasound images for medics operating in austere and urgent scenarios. Drexel experts collaborating on this effort include Information Science Assistant Professor Christopher MacLellan, PhD, Computer Science Associate Professor Edward Kim, PhD and Information Science Associate Professor Rosina Weber, PhD.
“DARPA is the premiere funding agency for artificial intelligence research. I am very excited that we were selected for such a competitive award with a team comprised entirely of Drexel researchers,” said Chris MacLellan, principal investigator of this initiative.
POCUS devices assist medical professionals to diagnose critical combat injuries. Though these portable ultrasound machines promise more accurate assessments and rapid results, they require costly and lengthy training modules to operate effectively. AI could help streamline POCUS processes for faster and less expensive deployment.
To successfully overcome POCUS AI challenges, Drexel’s expert research team proposed an innovative project titled SPARTACUS-X. “A major challenge for modern deep-learning systems is their massive data requirements. SPARTACUS-X will advance our understanding of how we can leverage human knowledge and existing materials to reduce data needs without compromising performance,” MacLellan explained.
SPARTACUS-X leverages existing POCUS sector information collected from Subject Matter Experts (SMEs) and ultrasound images to learn high-performance POCUS AI models for specific tasks with limited training data. Models created by SPARTACUS-X can explain their diagnoses, mitigating the consequences of misdiagnoses and accelerating POCUS AI adoption by military medics. Furthermore, SPARTACUS-X models improve accessibility by providing users with timely diagnoses and explanations on mobile devices.
“In order to train a machine to recognize ultrasound images, we are using a representation learning algorithm called sparse coding. Sparse coding has primarily been developed for computer vision, with successful applications in denoising, up-sampling, compression and object detection,” explains co-principal investigator Edward Kim. “Since sparse coding is a self-supervised learning algorithm, it does not require a large, labeled dataset to compute an efficient code.”
SPARTACUS-X improves existing POCUS AI by developing four vital innovations that address current challenges and limitations. These advancements include integrating Teachable AI to efficiently channel expert domain knowledge, implementing Sparse Coding to extract POCUS data from unlabeled images, leveraging gathered information to create Small-Data Classifiers that optimize POCUS accuracy with limited training data, and applying Knowledge-based XAI (Explainable Artificial Intelligence) to explain resulting diagnoses and support SME reasoning.
Drexel’s College of Computing & Informatics (CCI) contributes to theory and practice along dimensions that include technical, human, organizational, policy and societal considerations. This broad perspective positions the College to address the complex, multi-disciplinary problems that are increasingly common as society becomes more dependent on information technology. CCI boasts one of the oldest continually ALA-accredited library and information science programs in the US: the Library and Information Science major in the College’s Master of Science in Information degree program.