DB-ATRG: The Density and BI-RADS-Aware Triage and Automatic Report Generation System for Mammography

DB-ATRG fine-tunes MedGemma 1.5 on DMID to automate mammography report text and triage, achieving ROUGE-L 0.8012 and prioritizing BI-RADS 4–5 cases early.

• Why it matters: Radiologist shortages and FIFO queues can delay urgent or complex mammography interpretations, creating a need for risk-aware prioritization and faster report generation.
• What they did: The system uses QLoRA fine-tuning of a 4-billion parameter vision-language model, then applies a dual-phase triage that flags ACR Category D and ranks cases via a Cumulative Urgency Score.
• The result: In a simulated cohort of 100 cases, the optimized queue captured 100.0% of BI-RADS 4 and 5 malignancies within the first 20% of workload (vs 19.2% for FIFO) and reduced mean rank from 42.0 to 7.5.

AI summary of the abstract · may contain errors · not medical advice · read the paper.