30-Second Takeaway
- Generative and variance-based angiographic AI can cut intraprocedural dose by ~70–80% without sacrificing safety or image utility.
- LLM-based CT protocoling, with careful prompting, matches or exceeds radiologist performance, without needing fine-tuning.
- Low-dose CT and AI/DL workflows demand explicit correction and governance to avoid biased quantification and hidden cancer risk.
- Non-contrast, AI-enabled breast DWI and structured S modifiers in lung screening support richer, risk-focused reporting.
- LI-RADS 3 lesions show substantial short-term progression, favoring size- and liver-function–tailored follow-up intervals.
Week ending January 10, 2026
Radiology Grand Rounds: Practical AI, Dose Reduction, and Risk Stratification Across CT, MRI, and Interventional Practice
Generative AI DSA cuts intraoperative radiation dose by about two-thirds without compromising safety
In this randomized trial of 1068 surgical candidates, GenDSA-V2 reduced DSA radiation dose by roughly two-thirds versus standard protocols. Air kerma fell from 457.4 ± 407.4 mGy to 151.3 ± 125.1 mGy, and DAP decreased similarly, both with superiority P < 0.001. Operation time was non-inferior, differing by only 1.7 minutes, and intraoperative complication rates were similar between groups. The study spanned cerebral, thoracic, and hepatic interventions, supporting broad applicability for dose reduction in DSA-guided procedures. These data justify institutional evaluation of generative AI DSA as a practical strategy to lower patient and operator radiation burden.
GPT-4o, with optimized prompts, outperforms radiologists for abdominopelvic CT protocol selection
This retrospective study evaluated GPT-4o for automated protocoling of abdominal and pelvic CT scans in 1448 patients. With context-rich prompting, GPT-4o chose optimal protocols in 96.2% of cases versus 88.3% for radiologists (P < .001). Inappropriate protocol rates were similar between GPT-4o and humans, and fine-tuning conferred no additional benefit beyond prompting. Performance was consistent across trainees and attendings, suggesting LLM assistance could standardize protocol quality across experience levels. These findings support deployment of carefully governed LLM tools for CT protocoling rather than investing in model fine-tuning.
Higher trauma CT DLP is associated with increased incident cancer and cancer-related mortality
This statewide cohort included 2662 adult trauma patients without recent cancer, followed a median of 5.9 years. Patients underwent a median of three CT scans during the index admission, with median DLP of 1941 mGy·cm. New-onset cancer risk increased with CT dose, with an adjusted hazard ratio of 1.08 per 1000 mGy·cm increment in DLP. Cancer-related mortality was higher in patients exposed to DLP above 5000 mGy·cm, with adjusted hazard ratio 3.35. Results were consistent in a larger imputed cohort, reinforcing the need to justify and minimize CT exposure even in trauma settings.
DL on non-contrast breast DWI matches abbreviated contrast MRI performance and halves reading time
This four-center study developed a DWI-only deep learning model (DWI-DL) for breast cancer diagnosis in 2493 patients. Across internal, external, and prospective cohorts, DWI-DL achieved AUCs comparable to an abbreviated contrast-enhanced DL model. DWI-DL outperformed expert radiologists reading DWI alone, with higher AUCs across multiple cohorts. In a multireader study, an AI-guided selective sequence protocol was non-inferior to the full protocol while reducing interpretation time by 55.5%. These findings support non-contrast, AI-augmented breast MRI workflows that reduce gadolinium use and radiologist reading burden.
References
Numbered in order of appearance. Click any reference to view details.
Additional Reads
Optional additional studies from this edition.