Prompt design key for multimodal LLMs in femoral head osteonecrosis diagnosis: imaging plus text beats images alone

Pairing hip X-ray or MRI with radiology text descriptions boosted multimodal LLMs' detection of femoral head osteonecrosis: AUC 0.91 vs 0.55 for image-only, and staging reliability ICC 0.97 vs 0.51. Multi-image input offered no benefit.

• With added radiology descriptions, detection accuracy reached 0.88 (radiographs) and 0.85 (MRI), up from poor image-only performance.
• For differentiating early vs. late disease, ID prompts improved accuracy to 0.78 vs. 0.65 with images alone.
• No significant difference was found between commercial and open-source models (P=.83).

Automated summary

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