MIT-led team uses AI agents to clean up noisy PET scans
Multi-agent AI framework cuts PET scan noise 50% with no human help...
Deep Dive
A VLM- and LLM-driven multi-agent system for PET image denoising dynamically assesses image quality and lesion status, autonomously selects optimal denoising models and parameters, and enables closed-loop feedback with rollback. Tested on Siemens Biograph Vision Quadra PET/CT data at 1/20 and 1/50 low-dose settings, it achieved higher PSNR and SSIM than UNet, GAN, and DDPM baselines. These results show a closed-loop multi-agent framework can adapt PET denoising strategies to different image conditions.
Key Points
- Built by MIT, Ruhr-Universität Bochum, and collaborators, using VLMs for image quality assessment and LLMs for contextual reasoning.
- Tested on Siemens Biograph Vision Quadra with 1/20 and 1/50 dose settings, outperforming UNet, GAN, and DDPM baselines in PSNR/SSIM.
- Closed-loop multi-agent system autonomously selects denoisers, detects artifacts, and adapts parameters without expert input.
Why It Matters
Automates noisy PET scan cleanup, cutting radiologist workload and improving diagnostic reliability in low-dose imaging.