WAMS: Wavelet-Domain Adaptive Missing-Modality Synthesis for Arbitrary Missing-Modality MRI

Clinical brain MRI is often acquired with incomplete modality sets, making arbitrary missing-modality synthesis essential for robust tumor analysis. Most existing methods handle arbitrary inputs through VAE-style or shared latent representations, where available modalities are compressed into a compact space before reconstruction. Although effective in improving input flexibility, this compression-based paradigm weakens the direct modeling of image structures and limits the fidelity of synthesized modalities. We introduce Wavelet-domain Adaptive Missing-modality Synthesis (WAMS), which is the first framework to leverage wavelet transformation for arbitrary missing-modality synthesis. Instead of reconstructing MRI from a compressed latent space, WAMS decomposes multimodal MRI into low-frequency structural maps and high-frequency detail maps, where the low-frequency representation provides anatomical guidance for high-frequency detail generation. To adapt to different missing patterns, WAMS dynamically generates missing-state-aware priors according to the available modality. During synthesis, it completes only the unavailable modalities in the wavelet domain, while directly retaining the wavelet representations of already acquired modalities for inverse transformation, avoiding unnecessary re-generation of observed scans. Extensive experiments demonstrate that WAMS achieves superior synthesis quality, structural fidelity, and observed-modality preservation over existing methods.