Phantom-based synthetic renal PCASL data generation for overcoming data scarcity and enhancing deep learning-based segmentation.
Authors
Affiliations (10)
Affiliations (10)
- Computer Assisted Clinical Medicine, Medical Faculty Mannheim, Heidelberg University, Heidelberg, Germany.
- IdiSNA, Health Research Institute of Navarra, 31008, Pamplona, Spain.
- Institute of Smart Cities (ISC), Health Research Institute of Navarra, 31006, Pamplona, Spain.
- Department of Radiology, Clínica Universidad de Navarra, 31008, Pamplona, Spain.
- Department of Nephrology, Clínica Universidad de Navarra, 31008, Pamplona, Spain.
- Electrical Electronics and Communications Engineering, Public University of Navarre, 31006, Pamplona, Spain. [email protected].
- IdiSNA, Health Research Institute of Navarra, 31008, Pamplona, Spain. [email protected].
- Institute of Smart Cities (ISC), Health Research Institute of Navarra, 31006, Pamplona, Spain. [email protected].
- Mannheim Institute for Intelligent Systems in Medicine, Medical Faculty Mannheim, Heidelberg University, Heidelberg, Germany.
- Department of Radiation Oncology, Medical Faculty Mannheim, Heidelberg University, Heidelberg, Germany.
Abstract
We hypothesize that a 2D CycleGAN framework can generate anatomically realistic synthetic renal PCASL images from phantom to overcome data scarcity and improve deep learning-based segmentation models. A CycleGAN was implemented for style transfer from XCAT phantoms to PCASL. Datasets included 16 transplanted (TK) and 14 healthy (HK) subjects (3T Siemens Skyra). The synthetic dataset was used both to fine-tune existing models and to enable multi-class segmentation (cortex/medulla) in real images via a model trained exclusively on synthetic data, addressing the lack of labels. In HK, the mixed model (real + synthetic data) significantly improved segmentation (mean DSC increase:, p < 0.05). For isolated left HK, the synthetic-only model achieved the highest gain (p < 0.001). In TK, synthetic data alone was less effective (p > 0.05), but the mixed model maintained baseline performance (p < 0.05). Furthermore, training the Mask R-CNN exclusively on synthetic data enabled successful segmentation of the cortex (DSC: 0.659) and medulla in real images. Synthetic data enhances generalization and enables complex compartment segmentation. Our results demonstrate that GAN-based data augmentation significantly enhances segmentation in PCASL imaging, providing a robust and scalable solution for low-resource clinical settings.