ADLS: Alignment of discrete latent spaces for unsupervised cross-modality medical image translation

Unsupervised cross-modality medical image translation is essential when paired multi-modality data are unavailable. Traditional methods often fail to align domains effectively when trained on mini-batches containing structurally dissimilar slices from distant anatomical locations. We propose ADLS, a framework performing image translation by aligning representations in a shared discrete latent space. Our approach employs vector quantization with codeword stacking to enhance representational capacity and introduces a joint distribution matching loss aligning statistical dependencies between spatially adjacent codewords. We evaluated the proposed method on pelvic CBCT-to-CT and head PD-to-T2 translation tasks. ADLS outperformed state-of-the-art methods, with statistical tests confirming that the performance improvement is statistically significant. Source code is available at https://github.com/chenxu31/ADLS.