CycleGAN-derived synthetic renal PCASL images from phantoms boost deep learning segmentation

Synthetic PCASL images generated via CycleGAN from phantoms improved deep learning renal segmentation. A model trained only on synthetic data segmented cortex with Dice score 0.659 in real kidneys.

• A Mixed model (real + synthetic data) significantly improved segmentation in healthy kidneys (p<0.05), though exact Dice increase not reported in the abstract.
• When trained solely on synthetic images, Mask R-CNN achieved a mean Dice of 0.659 for cortex segmentation in real kidneys.
• Study included 16 transplanted (TK) and 14 healthy (HK) kidney subjects imaged at 3T.

Automated summary

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