Inconsistency of AI in Intracranial Aneurysm Detection with Varying Dose and Image Reconstruction
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Scanner-related changes in data quality are common in medical imaging, yet monitoring their impact on diagnostic AI performance remains challenging.
Purpose
To assess the consistency of an FDA-approved AI solution for intracranial aneurysm detection across changes in image data quality caused by dose and image reconstruction.
Methods
Consistency testing was performed using a head CT phantom designed for AI evaluation, replicating a patient with three intracranial aneurysms in the anterior communicating artery (ACoA), middle cerebral artery (MCA), and basilar artery (BA). The phantom was examined three times at 21 doses ranging from 0.47 to 20.09 mGy using iterative reconstruction (IR) and filtered back projection (FBP). Aneurysm labeling by an AI solution approved for triage and notification of intracranial aneurysms was analyzed. Five neuroradiologists evaluated all examinations for aneurysm visibility.
Results
The AI solution labeled 74.6% of ACoA, 92.9% of MCA, and 2.4% of BA aneurysms, while reader experiments yielded aneurysm visibility rates of 98.6%, 99.8%, and 95.4%, respectively. The AI demonstrated stable performance within the medium dose range but produced inconsistent results at doses below 8 mGy with IR, 7 mGy with FBP, and also above 14 mGy with FBP. Doses below 1 mGy with IR and 2 mGy with FBP led to a complete lack of AI response to any aneurysm. Readers noted at least one visible aneurysm in every case and reported 100% visibility for all aneurysms at doses above 2 mGy.
Conclusions
Diagnostic AI shows inconsistent performance when image data quality is not optimal and requires more stringent quality standards than radiologists. Standardized consistency testing reveals AI performance issues when these quality requirements are not achieved.
