DB-ATRG: Density and BI-RADS - Aware Triage and Automatic Report Generation for Mammography
| Advisor | Dr. Russell Jeter (Georgia State University) |
| Published At | [In Review] DB-ATRG: The Density and BI-RADS-Aware Triage and Automatic Report Generation System for Mammography (doi: 10.64898/2026.07.22.26358655) |
| Presented At |
|
| DOI / Paper | https://doi.org/10.64898/2026.07.22.26358655 |
Abstract
Timely review and accurate documentation of screening mammograms remain critical challenges in breast imaging due to high patient volumes, radiologist fatigue, and the elevated masking risk in dense breast tissue. We developed the Density and BI-RADS-Aware Triage and Report Generation (DB-ATRG) framework to automate diagnostic text generation and enable risk-based case prioritization, flagging scans with extremely dense breasts (ACR Category D) for supplemental screening.
We fine-tuned the 4-billion parameter MedGemma 1.5 vision-language model using QLoRA (Quantized Low-Rank Adaptation) on the combined DMID and VinDr-Mammo datasets. The model achieved an ACR Breast Density Accuracy of 0.7039, a ROUGE-L score of 0.8650, and a METEOR score of 0.9001 for clinically faithful diagnostic report generation.
Furthermore, to address queue delays for high-risk patients, we introduced a Cumulative Urgency Score simulation on a 100-case clinical cohort. The DB-ATRG prioritization captured 100% of high-risk malignancies (BI-RADS 4 and 5) within the first 20% of the reading workload (compared to only 40% under traditional first-in, first-out review) and shifted the mean rank of severe cases from 42.8 down to 3.
Conference Poster
