FlexiD-Fuse: Diffusion Fusion for Medical Imaging

• FlexiD-Fuse is a diffusion-based image fusion network that flexibly integrates multi-modal medical images using a unified architecture.
• It reformulates the fusion challenge as a maximum likelihood estimation problem, embedding an Expectation-Maximization procedure within the diffusion process.
• The network eliminates the need for modality-specific rules and separate weights, enabling end-to-end processing for varying numbers of input images in clinical settings.

Searching arXiv for the target paper and a small set of directly relevant diffusion/fusion references. FlexiD-Fuse is a diffusion-based image fusion network for multi-modal medical image fusion that is designed to accommodate flexible quantities of input modalities and can end-to-end process two-modal and tri-modal fusion under the same weight (Xu et al., 11 Sep 2025). The method addresses a stated limitation of existing fusion systems, namely that they are typically specialized to a fixed number of inputs and therefore cannot directly process varying input quantities in clinical settings. Its central reformulation converts diffusion fusion from a fixed-condition input problem into a maximum likelihood estimation problem based on the diffusion process and hierarchical Bayesian modeling, with an Expectation-Maximization (EM) procedure embedded into diffusion sampling so that fused images can be generated independently of the number of input images (Xu et al., 11 Sep 2025).

1. Problem setting and design objective

The immediate target of FlexiD-Fuse is medical image fusion across modalities such as MRI, CT, PET, and SPECT, where different source images provide complementary physiological and anatomical information (Xu et al., 11 Sep 2025). In the formulation described for prior fixed-input systems, traditional fusion networks are represented as input-number-specific functions such as $\mathrm{F}_{mf_2}$ and $\mathrm{F}_{mf_3}$, each trained for a specific modality count and therefore requiring separate architectures and weights.

FlexiD-Fuse replaces that regime with a generalized fusion function,

$$I_f = \mathrm{F}_{gf}(I_{mod_1}, \ldots, I_{mod_n}; \theta_g),$$

where $n$ is flexible and $\theta_g$ is shared (Xu et al., 11 Sep 2025). In the reported medical setting, inference is performed for $n=2$ or $n=3$, but the framework is presented as supporting arbitrary input quantities through a single unified architecture and weight setting. The paper further states that this is achieved without modality-specific fusion rules and without ground truth fusion images.

A common misconception in this problem area is that “flexible-input” fusion can be reduced to training multiple specialized models and selecting one at inference time. FlexiD-Fuse is defined against precisely that assumption: its claim is not merely multi-configuration compatibility, but end-to-end fusion under shared weights with dynamic adjustment of modality contributions during inference (Xu et al., 11 Sep 2025).

2. Architectural composition

The architecture combines