A new deep learning algorithm can detect intracranial hemorrhage on head CT scans similarly to an experienced radiologist. Researchers believe it could eventually help physicians treat patients with traumatic brain injuries and strokes.
Developed by a team from UC San Francisco and UC Berkeley, the AI algorithm spotted signs of hemorrhage on scans in only one second and yielded the “highest accuracy to date” for this type of clinical application, wrote co-author Jitendra Malik, PhD, with UC Berkeley, in the Oct. 21 study published in Proceedings of the National Academy of Sciences.
“Given the large number of people who suffer from traumatic brain injury every day and are rushed to the emergency department, this has very big clinical importance,” Malik said in a UCSF news article. “That convinced me to work on this problem."
Imaging utilization is on the rise and, as a consequence, radiologists are burdened with reading thousands of images a day; often they’re searching for tiny abnormalities that may indicate a larger problem—especially when diagnosing brain hemorrhage. AI, however, holds promise to dramatically reduce this workload and even improve an expert’s accuracy.
Seeking to do just that, Malik and her colleagues trained a fully convolutional neural network on 4,396 head CT scans performed at UC San Francisco and its affiliated hospitals. On the training images, the researchers manually labeled each small abnormality down to the pixel level, making the dataset more robust than many used to train such algorithms.
Compared to the performance of four American Board of Radiology-certified radiologist, the algorithm produced a receiver operating characteristic area under the curve of 0.991—beating two of the four experts.
The algorithm also detected those tiny, dangerous abnormalities missed by experts. It even found their location within the brain and classified them by subtype, affording radiologists key pieces of information to determine the best treatment for their patients.
“The hemorrhage can be tiny and still be significant,” Pratik Mukherjee, MD, PhD, professor of radiology at UCSF, added in the same statement. “That’s what makes a radiologist’s job so hard, and that’s why these things occasionally get missed. If a patient has an aneurysm, and it’s starting to bleed, and you send them home, they can die."
Going forward, the team intends to test their algorithm on CT scans from trauma centers across the U.S.