Upward Dense Connection Network With High-Dimensional Feature Fusion for PET-to-CT Image Conversion

Positron Emission Tomography and Computed Tomography are complementary imaging modalities in modern medical diagnostics: PET reveals tissue metabolism and functional activity, while CT provides precise anatomical structure information. The integration of the two enables comprehensive evaluation of both structural and functional characteristics of lesions. However, due to constraints such as radiation dose and medical cost, patients in clinical practice often undergo only a single-modality scan. Therefore, cross-modal image conversion between PET and CT has significant research and clinical application value. To address the common issues of insufficient feature fusion and redundant model design in existing methods, this study proposes a novel generative network for PET→CT image translation, termed the Upward Dense Connection Generative Adversarial Network. The core innovations are as follows: (1) Upward Dense Connection Mechanism: During the decoding phase, multi-level features from the encoder are densely aggregated across layers, enabling full propagation and fusion of high-dimensional semantic information. This enhances structural detail reconstruction and texture consistency in the synthesized CT images. (2) Single-Channel Input/Output Mechanism: By leveraging the inherent grayscale characteristics of medical images, the traditional three-channel input/output architecture is optimized to a single-channel structure. This reduces the number of parameters by approximately one-third, significantly improving model compactness and training stability. Quantitative and qualitative evaluations demonstrate that the proposed method outperforms U-Net and its variants in terms of PSNR, SSIM, and overall visual quality. The model shows superior performance in structural detail preservation, edge sharpness, and functional region mapping, along with better convergence stability and generalization capability.