FADFNet: A fine-tunable and adaptive decomposition-fusion network for cross-dataset low-dose CT and low-dose PET image reconstruction.
Authors
Affiliations (10)
Affiliations (10)
- College of Information Science and Electronic Engineering, Zhejiang University, Hangzhou, China. Electronic address: [email protected].
- School of Information Science and Engineering, Lanzhou University, China. Electronic address: [email protected].
- School of Computing and Data Science, The University of Hong Kong, Hong Kong, China. Electronic address: [email protected].
- Department of Radiology, The First Hospital of Lanzhou University, Lanzhou, China. Electronic address: [email protected].
- Department of Radiology, The First Affiliated Hospital of Zhejiang University School of Medicine, Hangzhou, China. Electronic address: [email protected].
- Polytechnic Institute, Zhejiang University, Hangzhou, China. Electronic address: [email protected].
- Clinic for Nuclear Medicine, Bern University Hospital, Inselspital, Bern, Switzerland. Electronic address: [email protected].
- School of Computing and Data Science, The University of Hong Kong, Hong Kong, China. Electronic address: [email protected].
- School of Information Science and Engineering, Lanzhou University, China. Electronic address: [email protected].
- College of Integrated Circuits, Zhejiang University, Hangzhou, China. Electronic address: [email protected].
Abstract
Low-dose computed tomography (LDCT) and low-dose positron emission tomography (LDPET) are pivotal for minimizing radiation risks in clinical practice, yet they inherently suffer from noise-induced image degradation. Although deep learning has advanced low-dose reconstruction, existing methods often lack adaptability across heterogeneous datasets and modalities due to severe domain shifts and rigid model architectures. To address these challenges, we propose FADFNet, a Fine-tunable and Adaptive Decomposition-Fusion Network designed for robust cross-domain reconstruction. Predicated on a frequency-aware paradigm, the proposed framework incorporates a Wavelet Transform-based Decomposition Module to explicitly disentangle the input into low-frequency structural components and high-frequency textural details. These components are processed by dual-branch experts equipped with a Context-Aware Spatial-Channel Modulation mechanism, which leverages structural priors to guide precise texture recovery. Subsequently, a Frequency Domain-based Feature Pyramid Fusion Module progressively integrates multi-scale features via global spectral attention to ensure comprehensive information synthesis. Furthermore, we introduce a parameter-efficient fine-tuning strategy that facilitates lightweight adaptation to unseen scanner domains by exclusively updating the modulation and fusion modules while freezing the pre-trained backbone. Extensive experiments on four LDCT datasets and four LDPET scenarios demonstrate the superiority of FADFNet over state-of-the-art methods. Quantitative results reveal that FADFNet achieves optimal signal fidelity and perceptual quality with significantly reduced computational overhead compared to generative approaches. Blind clinical evaluations further confirm that our method yields reconstruction quality comparable to full-dose reference images, substantiating its potential for reliable deployment in diverse clinical environments. The source code is available at FADFNet.