Dual-Domain Pretraining Foundation Model for CT Image Reconstruction

Deep learning (DL) has been widely and successfully used for low-dose CT (LDCT), sparse-view CT (SVCT), and limited-angle CT (LACT) image reconstructions, which aims to decrease X-ray radiation dose during CT scanning. However, most existing DL-based CT image reconstruction methods are typically developed in a task-specific manner and require the prior knowledge of degradation types and levels that can be very complicated in practice, thus limiting their application in real-world clinical scenarios. To address this issue, in this paper, we propose a dual-domain pretraining foundation model (CTPretrain) for multiple low-quality CT reconstruction tasks (i.e., LDCT, SVCT, LACT). Specifically, CTPretrain consists of a sinogram-domain restoration transformer (SRT) and an image-domain restoration transformer (IRT). First, the IRT module was pre-trained through a learnable Gaussian bandpass decomposition module for high/low-frequency separation, mining, and modulation, enabling adaptive weighting of informationrich frequency bands to address diverse degradation types and intensities. By employing a composite low-frequency degradation contrastive loss, positive sample pairs were brought closer while negative pairs were pushed apart in the feature space, thereby enhancing the model’s discriminative capability across degradation patterns. Subsequently, the frozen IRT module’s reconstruction outputs were utilized to effectively compensate for sparse projection data, which served as inputs to the SRT module to stabilize reconstruction performance in the projection domain. Experimental results demonstrate that our CTPretrain surpasses state-of-the-art methods on the hybrid dataset, achieving the capability to address multiple LDCT, SVCT and LACT tasks. Code and models will be available at https://github.com/LinY-ct/CTPretrain.