Multimodal LLM accurately segments soft tissue sarcomas on MRI in zero-shot setting

Gemini 3.0 achieved mean Dice 0.935 for zero-shot STS segmentation (radiologist inter-rater Dice 0.949). Tumor size from masks had ICC 0.976 vs 0.922 for direct LLM estimate (p=0.027), with bias -0.5 mm vs -5.0 mm. Segmentation-derived measurements are more reliable.

• Zero-shot segmentation by Gemini 3.0 produced masks with high reproducibility (Dice 0.981) and accuracy comparable to board-certified radiologists.
• Deriving diameter from automated masks yielded significantly higher agreement with manual measurement (ICC 0.976) than direct LLM estimation (ICC 0.922; p=0.027).
• Bland-Altman analysis showed a smaller mean bias for the segmentation-based approach (-0.5 mm) vs direct estimation (-5.0 mm).

Automated summary

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