Artificial Intelligence: The Ethics and Abilities – Diagnostic Imaging
AI Offers “Expert-Level” Detection of Brain Hemorrhage
Acute intracranial hemorrhage (ICH), sometimes referred to as a “brain bleed,” shares symptoms with several other neurological conditions. Today, emergency departments rely on CT scans to detect this life-threatening condition—and even the most experienced radiologists can sometimes miss the subtle signs of the condition on such lower resolution images. Now, researchers from the University of California, San Francisco and the University of California, Berkeley have demonstrated that a deep learning artificial intelligence (AI) algorithm can provide “expert-level” detection of brain hemorrhage in a new study published in the Proceedings of the National Academy of Sciences—not only performing at the same standard as expert radiologists but finding tiny brain bleeds that those experts overlooked.
The researchers used a single-stage, end-to-end, fully convolutional deep learning neural network in order to help identify what are usually very small abnormalities that must been detected on an image known for poor soft tissue contrast and low signal-to-noise issues. They trained the algorithm on a data set of over 4,000 CT exams where ICH abnormalities were manually highlighted at the pixel level.
While the authors said this data set was small, they argue that using this pixel-level supervision approach allowed for the kind of joint classification and segmentation to allow the network to better detect potential issues on the testing data. In addition, the authors used a technique called PatchFCN, where the researchers contextualized the training images with those that came both immediately before and after each one in the stack.
When the researchers compared the algorithm’s performance with four experienced, American Board of Radiology-certified radiologists on a test set of 200 head CT exams, they found that the algorithm was able to successfully detect ICH, achieving 100% sensitivity at specificity levels approaching 90%. Furthermore, the algorithm also identified some abnormalities missed by the expert reviewers – five positive cases were judged as negative by two of the four radiologists.
The researchers concluded that their algorithm could be used to augment radiologists’ abilities to detect ICH in the emergency department and plan to continue testing it in follow-on studies.