CTNEE-Net: A cascaded transformer network with edge enhancement for low-dose femoral head CT image denoising

Abstract In recent years, low-dose computed tomography (LDCT) image denoising techniques for the femoral head have become a prominent research focus in medical image processing because they maintain diagnostic accuracy while reducing radiation exposure. However, current methods still face challenges in preserving edge structures and facilitating collaborative multi-domain feature modeling. To address these challenges, this paper proposes an end-to-end cascaded transformer network with edge enhancement for LDCT image denoising (CTNEE-Net), which employs a multi-stage learnable unfolding architecture to synergistically optimize denoising and edge enhancement specifically for femoral head LDCT images. The network introduces a Learnable Edge Feature Extraction Module (LEFEM) that adaptively enhances bone edge details while suppressing noise. It also incorporates a Hybrid Transformer Block (HTB) that integrates frequency-domain attention and spatial self-attention to improve global context modeling. Furthermore, an Advanced Edge-Aware Fusion Module (AEAM) is proposed, which utilizes a gating mechanism to enable cross-domain adaptive fusion of edge features and deep semantic features, thereby enhancing the recovery of fine structures. Experimental results on 3 mm slice data from the Mayo dataset show that CTNEE-Net achieves PSNR and SSIM values of 34.05 dB and 0.9229, respectively. Subjective evaluations on clinical data also demonstrate its superiority over comparative methods. By effectively addressing the insufficiency of multi-domain feature synergy, this study offers significant advantages in preserving edge structures and enhancing denoising performance, thereby providing a solution that combines high performance and clinical practicality for LDCT image post-processing.