LLM-guided dual-decoder framework synthesizes multiphase contrast CT and improves liver lesion classification

Synthetic multiphase CT from noncontrast CT, generated by a large language model–guided dual-decoder framework, improved focal liver lesion classification accuracy from 71.65% to 85.04% (acquired CT 91.34%) in 86 patients.

• The framework uses a shared Swin Transformer encoder with dual decoders and Qwen3-8B-derived task prompts to synthesize arterial and portal-venous phases.
• Image quality metrics (PSNR, SSIM, MSE, PCC) were superior to those of Pix2pix and MedGAN in both whole-image and lesion-focused evaluations.
• External validation, clinically oriented safety assessment, and reader studies remain necessary before any clinical decision support or triage use.

Automated summary

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