Physics-informed deep learning trained on simulated data corrects motion artifacts in dynamic wrist CT

A physics-informed deep learning model trained entirely on simulated wrist 4DCT data effectively corrected motion artifacts, improving LIRS in cadaveric bone regions from 0.58±0.19 to 0.77±0.14 (p<0.01) and SSIM in simulated scans from 0.87 to 0.95.

• The simulation framework produced realistic motion artifacts comparable to real scans (p>0.05), providing paired clean/noisy data for training.
• The DL model, trained only on simulated data, generalised to cadaveric and in vivo scans, significantly improving image quality in bone and marker regions.

Automated summary

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