Which Generative AI Model is the Best Diagnostician?

Study evaluates LMMs for lung cancer CT scan diagnostics.

Models tested include GPT-4V, LLaVA-1.5, and BiomedCLIP.

GPT-4V had the highest agreement of 75 percent.

Accuracy for NCCN risk assessment was under 50 percent.

LLaVA-1.5 produced the most consistent descriptive results.

Models exhibited hallucinations and alignment issues.

Future improvements needed through collaboration with radiology professionals.