Deep Learning Detects Brain Hemorrhages with Accuracy Rivaling Experts

by Isha Salian

There’s a maxim in stroke treatment: “time is brain.”

It’s a reminder that during a stroke, human nervous tissue is lost at an alarming rate. From the onset of a brain injury and the start of medical treatment, every moment matters.

Researchers at UC Berkeley and UCSF School of Medicine are working on a deep learning model to reduce the time it takes to diagnose intracranial hemorrhage (bleeding in the skull) on a CT scan. With more than 80 million CT scans performed annually in the United States alone, AI could increase radiologists’ efficiency amidst an overwhelming volume of images.

“When someone gets into a car accident, or there’s a fall or other trauma that involves the head — a doctor will order a head CT scan,” said Weicheng Kuo, lead author on the paper, published this week in the leading scientific journal PNAS.

It’s critical to detect hemorrhages, even tiny ones. The analysis requires a high degree of focus by specially trained radiologists.

A neural network that assesses CT scans could reduce the burden on these specialists. Kuo and his collaborators expect a significant increase in radiologists’ productivity with their deep learning system, PatchFCN.

The researchers used NVIDIA V100 Tensor Core GPUs through Amazon Web Services for both the training and inference of their AI model, which segments the hemorrhage area, and identifies brain hemorrhages with 99 percent accuracy. The neural network also performs automated measurements of abnormalities on the CT scans, a  time-consuming step that radiologists manually perform today.

Every Minute Matters

Radiologists often have a large stack of scans to go through. Depending on how busy the day is, the turnaround time for reading a scan and reporting results to the emergency department can be half an hour or more.

If an abnormal case is at the bottom of the stack, the delay in diagnosis can adversely affect patients. AI can close this gap, processing a scan within a second, on average, using a single NVIDIA GPU for inference.

Trained on a dataset of more than 4,000 head CT scans from UCSF-affiliated hospitals, the PatchFCN performance rivals that of experienced radiologists, the study showed.

The volume of medical imaging studies is on the rise. Tools like PatchFCN could help radiologists manage larger workloads and boost their efficiency, Kuo says.

Looking Patch by Patch

Many convolutional neural networks analyze a whole image at once to come up with a result. And that makes sense: in the world of digital data, it’s often assumed that more information is better.

But, instead, the team found that splitting a CT scan into smaller patches improved the neural network’s results. They experimented with the patch size to achieve the best performance.

Neural networks can be set at different levels of recall, or sensitivity. The researchers believe that for this clinical application, the system should operate at the highest possible recall level, since the consequences of missing a brain hemorrhage could be catastrophic.

With this high-recall setting, the model had an average precision of 99 percent for detecting hemorrhages, the highest classification accuracy to date.

Rather than providing only a “yes” or “no” result, the neural network also provides a detailed tracing of each hemorrhage. The ability to highlight abnormalities directly on the image is essential for clinical use, because neurosurgeons must visually confirm the locations of hemorrhages on a head CT exam to judge the need and approach for surgical intervention.