Cross-Modality Deep Learning Denoising for Low-Dose μSPECT: Transfer of PET-Trained U‑Net and Diffusion Models.
Authors
Affiliations (3)
Affiliations (3)
- Medical Image and Signal Processing, Department of Electronics and Information Systems, Faculty of Engineering and Architecture, Ghent University, Ghent 9000, Belgium.
- J. Crayton Pruitt Family Department of Biomedical Engineering, University of Florida, Gainesville, Florida 32611, United States.
- Department of Radiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia, Pennsylvania 19104, United States.
Abstract
Reducing radiation dose in microsingle photon emission computed tomography (μSPECT) is essential to limit adverse biological effects in small animal studies, especially for longitudinal experiments. However, dose reduction increases noise and degrades image quality. While deep learning-based denoising (DL-DN) has shown promise for low-dose imaging, most prior work has focused on positron emission tomography (PET). DL-DN for μSPECT remains largely unexplored. In this study, we evaluate postreconstruction DL-DN for low-count μSPECT using two 2D architectures: a U-Net and a denoising diffusion probabilistic model (DDPM). To mitigate the limited availability of preclinical training data, we investigated transfer-learning from PET, a related molecular imaging modality, using models pretrained on both μPET (mouse) and clinical PET (human) data. These models were evaluated in a zero-shot setting and further adapted to the μSPECT domain via transfer-learning. Their performance was compared to models trained from scratch using varying amounts of μSPECT data (4-32 volumes). The data set consisted of in vivo mouse scans acquired with two tracers at 10%, 25%, and 50% of standard counts. Both architectures achieved meaningful noise reduction and improved quantitative agreement with standard-count references. Zero-shot PET-pretrained models demonstrated cross-modality generalization, reducing root mean squared error (RMSE) by more than 25% at the lowest count level, although residual artifacts remained due to differences between the PET source and μSPECT target domains. Transfer-learning consistently improved performance, yielding additional RMSE reductions exceeding 5% relative to zero-shot models. Its advantage over scratch-training decreased with increasing data set size, with maximal RMSE gains dropping from over 10% for small training sets to below 3% for larger data sets. Overall, the U-Net provided more robust performance with lower computational cost. These results indicate that DL-DN can support dose reduction in μSPECT, and that transfer-learning from PET offers a practical approach to address the small data set sizes common for preclinical settings.