Diffusive deformation priors for medical image synthesis

構造骨格

The source work combines diffusion with deformation-aware structure so that synthesis respects anatomical change rather than only appearance matching.

The key structural question is whether the target domain change can be represented as a path through admissible deformations. A synthesis model that only learns endpoint appearance can generate plausible images while breaking the underlying anatomy. The ISOM layer isolates the deformation path as the object to transfer: generation should preserve the structure that a downstream measurement or clinician would treat as real.

物理概念／数学的対象

The transferable structure is a transport process constrained by geometry: useful samples move through a deformation field instead of hopping between unrelated image states.

AIターゲット問題

Target multimodal generative systems where outputs must preserve latent structure, such as medical synthesis, simulation-to-real adaptation, or representation alignment across sensing modalities.

変数・演算子・目的のマッピング

• Deformation field -> latent transport map between source and target conditions
• Diffusion trajectory -> uncertainty-aware refinement path
• Anatomical consistency -> structure-preserving constraint on generation

なぜこれがうまくいくのか

Many generative pipelines learn appearance transfer without a disciplined account of what should stay anchored. A deformation-aware transport view gives the model a place to store what moves, what stays, and what remains uncertain.

Diffusion already gives a staged refinement process, but without an explicit deformation prior it may spend capacity correcting errors it created earlier. Putting transport ahead of denoising gives the model a structured intermediate variable: where the object moves, what remains fixed, and where uncertainty should stay visible. That can make the generated sample more auditable than a single opaque image translation.

なぜ失敗するのか

If the target domain change is not well-approximated by a transport process, the prior becomes restrictive. Poor deformation estimates can also inject structural errors that diffusion then amplifies.

最小の反証可能な実験

On a paired multimodal benchmark, compare a baseline diffusion model against one that explicitly predicts a latent deformation transport prior before denoising. Measure structural consistency, uncertainty calibration, and downstream task utility. Reject the brief if the transport prior improves visuals but not structural metrics or calibration.

The decisive ablation is not visual preference. Compare downstream segmentation, landmark consistency, and uncertainty calibration under the same diffusion backbone with and without the deformation transport variable. Reject the transfer if the added prior only improves surface realism while anatomical measurements, calibration error, or task utility remain unchanged.