Tuesday, December 3 | 9:00 a.m.-9:30 a.m. | T2-SPNR-4 | Learning Center
How useful is AI for detecting critical brain imaging findings on MRI? It's on par with neuroradiologist readers -- and thus could prompt appropriate further imaging and reduce callbacks, according to a poster to be highlighted Tuesday morning.
Kaining Sheng, MD, of Rigshospitalet in Copenhagen, Denmark, will share results from a study he conducted with colleagues that found that "using AI as a human support to implement a real-time adaptive MR scan acquisition workflow … can potentially reduce scan recalls."
Sheng's team assessed the diagnostic accuracy of an AI model called Apollo (Cerebriu) compared with that of three consultant neuroradiologists and three MRI technologists for detecting critical brain imaging findings via a study that included 414 patients, 65 of whom had confirmed brain infarcts, 65 of whom had intracranial hemorrhages, and 65 of which had intracranial tumors. All study participants underwent a three-sequence protocol (DWI, SWI/T2*, and T2-FLAIR).
The group found the following:
- Neuroradiologists detected infarct, hemorrhage, and tumors with a sensitivity of 88%, 85%, and 72%, respectively, and a specificity of 99%, 98%, and 97%, respectively.
- MRI technologists detected infarct, hemorrhage, and tumors with a sensitivity of 92%, 69%, and 35%, respectively, and a specificity of 90%, 95%, and 97%, respectively.
- Apollo detected infarcts, hemorrhages, and tumors with a sensitivity of 94%, 83%, and 71%, respectively, and a specificity of 86%, 84%, and 60%, respectively.
"Given an abbreviated three-sequence protocol set, Apollo [had] on par sensitivity with neuroradiologists and better sensitivity than MR technologists in recognizing critical findings in need of specialized scan sequences, thus providing a framework for relevant and configurable on-the-fly scan adaptation," Sheng and colleagues noted.
Visit this session to get all of the details.
