End-to-End Multiagent AI System Outperforms Initial Reads for Glioma BT-RADS Scoring

In a retrospective study of 492 posttreatment glioma MRI exams, an automated AI system achieved 76.0% accuracy for Brain Tumor Reporting and Data System (BT-RADS) scoring, compared with 57.5% for initial assessments (absolute improvement +18.5 percentage points, P<.001).

• The system combined a large language model for clinical variable extraction and a convolutional neural network for volumetric analysis.
• Context-dependent BT-RADS categories showed high sensitivity (e.g., BT-1b 100%, BT-1a 92.7%), while BT-4 positive predictive value was 92.9%.
• Cohen’s kappa between the system and expert reference standard was 0.71.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.