HCLD: Physics-aware and metadata-adaptive latent diffusion for high-fidelity medical image compression

High-fidelity medical image compression confronts a dilemma between the computational efficiency required for large-scale storage and the rigorous texture preservation demanded by clinical diagnosis, further complicated by the unique physical noise characteristics of medical imaging. To address this, we propose Hybrid-Conditioned Latent Diffusion (HCLD), a Physics-Aware and Metadata-Adaptive Latent Diffusion model. Specifically, we employ VQ-VAE-2 to shift the generative process into a compact latent space, significantly reducing computational costs. Addressing the mismatch of standard Gaussian assumptions, a Physics-Aware Hybrid Noise Injection mechanism couples Poisson and Gaussian statistics to actively preserve signal-dependent tissue textures. Furthermore, a Metadata-Adaptive Cross-Attention Mechanism (MACA) integrates clinical priors (e.g., scanning parameters) to guide anatomically accurate reconstruction. Extensive experiments on TCGA-LUAD and fastMRI demonstrate state-of-the-art performance (PSNR: 43.48 dB and 37.71 dB) with superior perceptual fidelity. Clinical validation via a downstream diagnosis task further confirms its reliability, achieving 91.16% accuracy comparable to uncompressed data. Our code is available at https://github.com/Shadows28512/HCLD-Medical-Compression.