Shared Representation Learning for Joint CT Reconstruction and Anatomical Segmentation

Denoising-based CT reconstruction methods can suppress high-frequency textures that are relevant for subtle lesion visibility. Motivated by hybrid convolution–attention designs such as CTLformer, this paper proposes a frequency-constrained denoising framework that preserves diagnostically relevant textures while reducing noise. The method introduces a dual-domain loss combining spatial fidelity with frequency-band constraints computed using discrete cosine transform representations. Evaluations on 52,000 paired slices from two low-dose CT datasets show that, relative to CNN-only and attention-only baselines, the proposed approach increases PSNR by 0.7–1.1 dB while maintaining higher high-frequency energy consistency. Reader-oriented texture metrics also improve by 8%–14% in regions with fine structural patterns.