Artificial Intelligence for Imaging of Brain Emergencies
Pratik Mukherjee, MD, PhD
Professor of Radiology and Biomedical Imaging, Bioengineering, UC San Francisco
Every 28 seconds, an American suffers a catastrophic neurologic emergency, most commonly stroke or traumatic brain injury (TBI). These injuries affect 15 million Americans every year and account for 7% of annual health care costs. Moreover, irreversible brain damage can begin to occur within minutes of injury, and immediate diagnosis and evaluation options have not been available to aid this immediate critical window.
Computed Tomography (CT) scanning
CT scans take a series of X-ray images to recreate the image of bones, blood vessels, organs, and other tissues inside the body and is frequently used to identify internal injuries and indications of disease. Currently, the scans are subject to radiologist interpretation and "grading" which is subjective and time consuming.
Artificial Intelligence (AI)
Artificial Intelligence is the employment of computer/machine-based technology simulations of intelligent human behavior to accomplish tasks. These often include tasks such as visual perception, decision-making, translation, and image recognition. AI technologies are created through many iterations of training a program to recognize input and perform tasks, making it more precise as it receives more data.
Advances in image recognition, such as those employed by Google and Facebook to recognize faces in images has yet to be applied in the same manner to medical imaging. Automated recognition of injury in brain images could improve rapid detection of emergencies as well as reduce physicians’ time constraints and subjectivity common in these diagnoses. The team successfully applied state-of-the-art AI to automate CT scan evaluation, and after training the AI with over 100,000 CT scans, they achieved detection of intracranial bleeding with greater than 99% accuracy. This is equivalent to the performance of board-certified radiologists.
Additionally, they accomplished integration of this technology into a cloud-based platform, which can be used with several different CT scanning devices and protocols at many hospitals. This allows medical professionals to use it anywhere in the world, and update a catalog of potential clinical biomarkers of neurologic injuries. This tool holds the potential to expedite treatment of irreversible damage and reduce subsequent long-term disability and death.
This project also recognizes the particular need for clinically effective, cloud-based, automated image analysis in underserved areas of the U.S. and developing world, as TBI and stroke disproportionately affect underprivileged populations. Provision of CT scanning equipment alone is not enough to meet the needs of these populations, as they are often still deficient in enough qualified radiologists to interpret the patient CT scans. Once approved for commercial use, the research team aims to make this technology available to meet the needs of Californians as well as patients around the world without rapid access to qualified diagnostic professionals.
To further test and apply their technology to integrative patient care, the team is working with a commercial partner to receive FDA approval for this device as a diagnostic test. Once approved, they will implement this technology with a focus on measurement of patient outcomes and application as the first quantitative biomarker of brain injury in clinical trials.
Research Team and Collaborators
University of California, San Francisco
- Arash Afshinnik, MD
- Claude Hemphill, MD
- Nerissa Ko, MD
- Geoff Manley, MD, PhD
- Pratik Mukherjee, MD, PhD
- Esther Yuh, MD, PhD
University of California, Berkeley
- Jitendra Malik, PhD