MFG Sampling: Solving Inverse Problems in Multi-level High-Frequency Guidance via Diffusion Models

Deep learning is widely applied in medical imaging, but models trained on one domain often underperform on another due to distribution shifts such as color and resolution differences across hospitals and equipment. Consequently, domain adaptation is essential to adjust target domain characteristics while preserving critical pathological features and patient-specific details. We propose a Fourier-based approach that retains source domain high-frequency components and adapts low-frequency content via a diffusion-model-based inverse problem. Rather than using fixed thresholds, we formulate a linear multi-level frequency extraction and guide sampling with our Multi-Level High-Frequency Guidance Sampling (MFG Sampling). This unsupervised method requires no paired data, offers noise robustness through frequency-based guidance, and can concurrently address sub-tasks such as super-resolution and deblurring. Classification experiments on a fundus dataset validate its effectiveness in domain adaptation.