Radiologist Perceptions of an AI Tool for Intracranial Hemorrhage Detection in Teleradiology: Cross-Sectional Survey Study.

Authors

Abstract

Artificial intelligence (AI) detection tools for intracranial hemorrhage (ICH) are increasingly integrated into radiology workflows. In real-world practice, perceived utility depends not only on diagnostic performance but also on workflow fit, false positive burden, and how clinicians interpret and act on AI outputs. This study aimed to characterize radiologists' perceptions of a Food and Drug Administration (FDA)-cleared ICH AI detection tool in a national teleradiology network, including perceived reliability, false positive burden, workflow impact, medicolegal concerns, and self-reported behaviors during routine use. We conducted an anonymous cross-sectional survey of radiologists in a national teleradiology practice who had access to an FDA-cleared ICH AI overlay during noncontrast head computed tomography interpretation. Survey items used a 5-point Likert scale. Results are summarized as agreement proportions ("agree" or "strongly agree") with 95% CIs. We compared neuroradiologists with non-neuroradiologists using Fisher exact tests. One primary end point was prespecified: agreement that time spent reviewing examinations with false positive AI alerts outweighed the benefits. Remaining subgroup comparisons were treated as exploratory, with false discovery rate control using the Benjamini-Hochberg procedure. A total of 65 radiologists responded, including 23 (35.4%) neuroradiologists and 42 (64.6%) non-neuroradiologists. Only 18.5% (12/65; 95% CI 10.9%-29.6%) agreed that false-positive alerts were infrequent enough to be acceptable. Agreement that the AI correctly identified most ICH cases was 32.3% (21/65; 95% CI 22.2%-44.4%), and agreement that the AI rarely missed clinically important hemorrhages was 43.1% (28/65; 95% CI 31.8%-55.2%). Trust in AI output was conditional: 50.8% (33/65; 95% CI 38.9%-62.5%) reported trusting the AI when it agreed with their interpretation, whereas 3.1% (2/65; 95% CI 0.8%-10.5%) reported trusting it when it conflicted with their interpretation. Only 10.8% (7/65; 95% CI 5.3%-20.6%) reported reduced overall interpretation time, whereas 33.8% (22/65; 95% CI 23.5%-46.0%) agreed that time spent reviewing false-positive alerts outweighed the benefits. Self-reported reduced scrutiny after an AI-negative result was uncommon (4/65, 6.2%; 95% CI 2.4%-14.8%). In subgroup analysis, neuroradiologists more often endorsed the primary end point than non-neuroradiologists (12/23, 52.2% vs 10/42, 23.8%; unadjusted P=.03), but no exploratory subgroup differences remained statistically significant after false discovery rate correction. Free-text responses emphasized artifact- and calcification-driven false positives, delayed or inconsistent AI availability, consultation burden, and medicolegal concerns. In this national teleradiology setting, radiologists reported substantial false positive burden, limited perceived time savings, and strongly conditional trust in an FDA-cleared ICH AI detection tool. Self-reported reduced scrutiny after negative AI outputs was uncommon but present in a minority of cases. These findings support the importance of specificity, interpretability, latency, and workflow-aware implementation when deploying radiology AI tools in practice.

Tags

Topics