Multi-Contrast Fusion Module (MCFM)

• Multi-Contrast Fusion Module (MCFM) is a design abstraction that integrates different contrast conditions and feature streams to enhance imaging clarity and maintain target specificity.
• It is applied across fetal ultrasound, MRI, and CT, employing techniques like branch-level attention and frequency-spatial fusion to manage complementary information.
• MCFM’s flexible, multi-stage architecture enables early shallow fusion and selective integration to address challenges such as misalignment and modality contamination.

Multi-Contrast Fusion Module (MCFM) denotes a class of feature-fusion mechanisms that integrate information from multiple contrast conditions, modalities, phases, or feature streams into a target representation. The term is used explicitly for an early-stage attention module in fetal ultrasound classification (Zhu et al., 13 Aug 2025), but in MRI and related medical-imaging literature the same functional role is often implemented under different names, including hierarchical fusion modules, multi-stage integration modules, cross-scale feature fusion modules, cross-modal selective fusion, iterative fusion modules, and frequency-spatial fusion blocks (Zhang et al., 2023, Feng et al., 2021, Yang et al., 2024, Chen et al., 2024, Huang et al., 2024, Li et al., 5 Dec 2025). Across these formulations, the common objective is to exploit complementary information while preserving target specificity and avoiding redundancy or modality contamination.

1. Terminology, scope, and usage

The phrase “multi-contrast” is not used uniformly across the literature. In fetal ultrasound, it refers to multiple intensity-adjusted copies of the same grayscale image, generated under predefined contrast settings and fused at the shallow-feature stage (Zhu et al., 13 Aug 2025). In multi-contrast MR reconstruction, it refers to a paired setting in which an undersampled target modality is reconstructed with guidance from a high-quality auxiliary modality acquired in the same session; the relationship is explicitly described as existing in both the spatial/image domain and the frequency/$k$-space domain (Li et al., 5 Dec 2025). In multi-phase contrast-enhanced CT, the analogous notion is temporal rather than modal, with phase set $P=\{\text{NC}, \text{A}, \text{V}, \text{D}\}$ and optional delayed phase (Huang et al., 2024).

This terminological variation matters because an MCFM may operate on raw images, encoded feature maps, temporal phases, or branch-specific latent tensors. It is therefore better understood as a design abstraction than as a single canonical block.

Context	What “multi-contrast” denotes	Functional MCFM form
Prenatal ultrasound	Contrast-adjusted copies of one grayscale image	Early branch-attention fusion
Multi-contrast MRI reconstruction	Target modality plus auxiliary/reference modality	Cross-modal feature fusion
Multi-phase contrast-enhanced CT	NC, A, V, and optional D phases	Iterative temporal fusion

A second terminological point is that many papers do not explicitly name a block “MCFM.” The breast CE-MRI synthesis paper states that its hierarchical fusion module, weighted difference module, and multi-sequence attention module are the functional equivalent of an MCFM (Zhang et al., 2023). Similar statements appear for SANet, ECFNet, and FASR-Net, where the effective fusion mechanism is distributed across stage-wise attention, cross-scale alignment, or progressive hierarchical fusion rather than encapsulated in one separately named unit (Feng et al., 2021, Yang et al., 2024, Liu et al., 2022).

2. Canonical computational forms

A minimal abstract description of an MCFM-like operator is given in the breast MRI synthesis setting, where per-sequence encoders produce latent features and a fusion operator combines them: $$F_i = E_i(X_i), \qquad F_{\text{fused}} = \phi(F_{T1}, F_{DWI,b0}, F_{DWI,b150}, F_{DWI,b800}, F_{DWI,b1500}).$$ In that paper, $\phi(\cdot)$ is not a single block; it is implemented hierarchically and includes learned DWI difference modeling, concatenation/fusion of multi-sequence features, channel attention refinement, and multi-scale integration (Zhang et al., 2023).

The explicitly named ultrasound MCFM uses branch-level scalar attention over contrast-conditioned shallow features. A faithful formalization given in the technical summary is: $$z^{(k)} = \frac{1}{C \times H' \times W'} \sum_{p=1}^{C} \sum_{i=1}^{H'} \sum_{j=1}^{W'} F^{(k)}_{pij},$$

$$a^{(k)} = \sigma(g(z^{(k)})), \qquad \tilde{F}^{(k)} = a^{(k)} F^{(k)},$$

$$F_{\text{fused}} = \operatorname{Concat}\left(\tilde{F}^{(1)}, \tilde{F}^{(2)}, \dots, \tilde{F}^{(K)}\right).$$

Here the salient property is that attention is computed per contrast branch, not per channel or per spatial location (Zhu et al., 13 Aug 2025).

A different canonical form appears in temporal multi-phase fusion. LIDIA defines per-phase feature extraction

$$\bm{h}^{p} = \mathcal{F}_p(\bm{x}^{p}) \in \mathbf{R}^C, \quad p \in P,$$

and iterative fusion

$$\bm{h}^{fuse}_{k+1} = \mathcal{F}_{conv}(\mathrm{concat}(\bm{h}^{fuse}_{k}, \bm{h})) \in \mathbf{R}^C,$$

with initialization $\bm{h}^{fuse}_1 = \bm{h}^{\text{NC}}$ and stopping at $P=\{\text{NC}, \text{A}, \text{V}, \text{D}\}$0 or $P=\{\text{NC}, \text{A}, \text{V}, \text{D}\}$1 depending on the availability of delayed phase (Huang et al., 2024). This formulation is explicitly order-dependent, reflecting the temporal progression of enhancement.

Taken together, these examples show that MCFM-like modules are typically feature-level rather than decision-level mechanisms. They may be single-shot, stage-wise, cross-scale, branch-selective, or iterative, but they are unified by the requirement that complementary information be transferred in a controlled rather than indiscriminate manner.

3. Explicit MCFM in fetal ultrasound classification

The paper “Multi-Contrast Fusion Module: An attention mechanism integrating multi-contrast features for fetal torso plane classification” introduces the term MCFM directly and gives the clearest standalone realization of the concept (Zhu et al., 13 Aug 2025). The task is a 4-class classification problem on fetal ultrasound images with classes 0–3 corresponding to transverse kidney, sagittal kidney, sagittal spine, and transverse abdomen. The paper motivates MCFM by the observation that fetal ultrasound images often exhibit low contrast, weak or unclear textures, blurred anatomical boundaries, high noise, low signal-to-noise ratio, and inconsistent image quality across cases.

The module is restricted to the lower layers of the network. Each image is resized to $P=\{\text{NC}, \text{A}, \text{V}, \text{D}\}$2, converted to single-channel grayscale, and normalized to $P=\{\text{NC}, \text{A}, \text{V}, \text{D}\}$3. Multiple contrast-adjusted versions are then created. The paper contains an explicit inconsistency: the main method description refers to three contrast-enhanced image sequences corresponding to contrast levels 1, 2, and 3, whereas the implementation details report four contrast levels $P=\{\text{NC}, \text{A}, \text{V}, \text{D}\}$4 and the experimental system appears to use four contrast branches (Zhu et al., 13 Aug 2025). The same source also notes a qualitative visualization under $P=\{\text{NC}, \text{A}, \text{V}, \text{D}\}$5, but treats that figure as illustrative rather than necessarily identical to the training configuration.

Each contrast-adjusted image is processed by a shallow convolutional branch with Mish activation, and each branch is said to produce three feature maps. Global Cross-Channel Pooling (GCCP) compresses each branch into a single scalar, and a neural network followed by Sigmoid produces one attention weight per branch. The weights are independent sigmoid gates rather than a softmax normalized across branches. Weighted branch features are concatenated and passed to a standard backbone; the reported integrations are with ResNet18 and ResNet34 (Zhu et al., 13 Aug 2025).

This architecture differs from standard SE-like channel attention in two respects. First, the attention unit is the contrast-conditioned branch rather than a channel in a single feature tensor. Second, the module operates before deep feature extraction, on shallow features derived directly from contrast-adjusted image inputs. The paper repeatedly describes this as a lightweight design. It also contains a numerical inconsistency in parameter overhead: the parameter table indicates ResNet18 increases from 11.18M to 11.21M and ResNet34 from 21.29M to 21.32M, implying about $P=\{\text{NC}, \text{A}, \text{V}, \text{D}\}$6M parameters, while the ablation text says approximately $P=\{\text{NC}, \text{A}, \text{V}, \text{D}\}$7 million parameters were added (Zhu et al., 13 Aug 2025).

4. MCFM-like mechanisms in multi-contrast MRI

In MRI, MCFM-like modules are usually embedded in larger reconstruction, synthesis, or super-resolution systems rather than named explicitly. A central distinction is that MRI fusion frequently spans both spatial/image and frequency/$P=\{\text{NC}, \text{A}, \text{V}, \text{D}\}$8-space domains. UniFS states this directly: the relation between contrasts exists in the spatial/image domain through similar structures, edges, and tissue boundaries, and in the frequency/$P=\{\text{NC}, \text{A}, \text{V}, \text{D}\}$9-space domain through shared high-frequency structural content but contrast-specific low-frequency intensity/style distributions (Li et al., 5 Dec 2025). Its unified model integrates a Cross-Modal Frequency Fusion module, an Adaptive Mask-Based Prompt Learning module, and a Dual-Branch Complementary Refinement module in order to handle multiple $$F_i = E_i(X_i), \qquad F_{\text{fused}} = \phi(F_{T1}, F_{DWI,b0}, F_{DWI,b150}, F_{DWI,b800}, F_{DWI,b1500}).$$0-space undersampling patterns without retraining.

FSMNet presents a closely related dual-domain design. Its Frequency-Spatial Feature Extraction module uses a frequency branch to capture global dependency with an image-size receptive field and a spatial branch to extract local features. Cross-Modal Selective fusion then selectively incorporates auxiliary frequency and spatial features into the corresponding target branches, and Frequency-Spatial fusion integrates the enhanced global and local target features into a comprehensive representation (Chen et al., 2024). This staged decomposition is significant because it separates cross-modal fusion from cross-domain fusion.

Stage-wise and cross-stage fusion appear prominently in multi-contrast MR super-resolution. SANet fuses auxiliary and target features at each stage through a separable attention compensation module $$F_i = E_i(X_i), \qquad F_{\text{fused}} = \phi(F_{T1}, F_{DWI,b0}, F_{DWI,b150}, F_{DWI,b800}, F_{DWI,b1500}).$$1, then applies a multi-stage integration module $$F_i = E_i(X_i), \qquad F_{\text{fused}} = \phi(F_{T1}, F_{DWI,b0}, F_{DWI,b150}, F_{DWI,b800}, F_{DWI,b1500}).$$2 to the full stack of intermediate features,

$$F_i = E_i(X_i), \qquad F_{\text{fused}} = \phi(F_{T1}, F_{DWI,b0}, F_{DWI,b150}, F_{DWI,b800}, F_{DWI,b1500}).$$3

to model inter-stage dependencies and obtain a holistic fused representation $$F_i = E_i(X_i), \qquad F_{\text{fused}} = \phi(F_{T1}, F_{DWI,b0}, F_{DWI,b150}, F_{DWI,b800}, F_{DWI,b1500}).$$4 for final reconstruction (Feng et al., 2021). MINet is built on the same premise: the multi-stage integration module matches each representation with all other features and integrates them in terms of their similarities (Feng et al., 2021).

Cross-scale variants push the same idea further. ECFNet uses a Cross-scale Feature Fusion Module (CFFM) to align and fuse target features across scales by deformable convolution and then applies a dual cross-attention transformer with reference features to generate coarse-to-fine textures $$F_i = E_i(X_i), \qquad F_{\text{fused}} = \phi(F_{T1}, F_{DWI,b0}, F_{DWI,b150}, F_{DWI,b800}, F_{DWI,b1500}).$$5; the Texture Transfer Module and Structure Information Collaboration Module subsequently inject texture and structure priors during decoding (Yang et al., 2024). FASR-Net separates alignment and fusion more explicitly: Flexible Alignment first addresses foreground-scale and patch-size mismatch through Single-Multi Pyramid Alignment and Multi-Multi Pyramid Alignment, then Cross-Hierarchical Progressive Fusion fuses aligned features progressively across scales with confidence-aware weighting (Liu et al., 2022).

The breast CE-MRI synthesis paper provides another distinct MRI interpretation of MCFM. Its effective fusion mechanism combines hierarchical multi-scale fusion, a weighted difference module specialized for DWI inputs and inspired by the ADC formulation, and a multi-sequence attention module that performs channel-wise attention over concatenated multi-sequence features. The inputs are $$F_i = E_i(X_i), \qquad F_{\text{fused}} = \phi(F_{T1}, F_{DWI,b0}, F_{DWI,b150}, F_{DWI,b800}, F_{DWI,b1500}).$$6, $$F_i = E_i(X_i), \qquad F_{\text{fused}} = \phi(F_{T1}, F_{DWI,b0}, F_{DWI,b150}, F_{DWI,b800}, F_{DWI,b1500}).$$7, $$F_i = E_i(X_i), \qquad F_{\text{fused}} = \phi(F_{T1}, F_{DWI,b0}, F_{DWI,b150}, F_{DWI,b800}, F_{DWI,b1500}).$$8, $$F_i = E_i(X_i), \qquad F_{\text{fused}} = \phi(F_{T1}, F_{DWI,b0}, F_{DWI,b150}, F_{DWI,b800}, F_{DWI,b1500}).$$9, and $\phi(\cdot)$0, and the target is contrast-enhanced T1-weighted MRI (Zhang et al., 2023).

These MRI examples show that an MCFM in practice may be frequency-spatial, cross-modal selective, stage-wise recurrent, or cross-scale and alignment-aware. What remains constant is the attempt to transfer complementary anatomical or spectral information without losing the target contrast identity.

5. Temporal, multimodal, and refiner-based extensions

The MCFM idea extends naturally beyond paired MRI contrasts. In multi-phase contrast-enhanced CT, LIDIA inserts its Iterative Fusion Module as the first layer of the encoder within an nnU-Net/Mask2Former-style framework. Each phase passes through a context-specific block consisting of two convolutional blocks with 3D convolutions, instance normalization, and LeakyReLU, after which phases are fused in the fixed temporal order NC $\phi(\cdot)$1 A $\phi(\cdot)$2 V $\phi(\cdot)$3 D (Huang et al., 2024). The delayed phase may be absent, so the module is explicitly designed for variable numbers of available phases, although the concrete setting is primarily 3-phase or 4-phase fusion rather than arbitrary missing-modality combinations.

A broader multimodal reformulation appears in Refiner Fusion Network. ReFNet is not an MCFM in name, but it is directly relevant as a pluggable fusion-plus-refinement architecture. A base fusion backbone $\phi(\cdot)$4 forms a fused embedding,

$\phi(\cdot)$5

and modality-specific decoder heads produce

$\phi(\cdot)$6

which are regularized against modality features by cosine similarity. The total training objective combines downstream loss, refiner loss, and Multi-Similarity contrastive loss on the fused embedding (Sankaran et al., 2021). This design reframes fusion as a partially invertible representation that must remain informative for each modality.

CoCoNet provides a different adjacent formulation. It targets infrared-visible image fusion and medical image fusion such as MRI-PET and MRI-SPECT through a coupled contrastive constraint and a multi-level attention module. The coupled contrastive objective pulls the fused image’s foreground target/background detail representation toward the infrared/visible source and pushes it away from the visible/infrared source, while the multi-level attention module is introduced to learn rich hierarchical feature representation and comprehensively transfer features in the fusion process (Liu et al., 2022). Although not labeled MCFM, the module addresses two canonical MCFM objectives: preserving complementary information and preventing information degeneration.

Analogous design patterns also appear in image-fusion systems that the source material explicitly describes as “analogous but not equivalent” to true MCFM. MATCNN’s MSFM combines multi-scale local feature extraction with Transformer-based global supervision through GFEM, and NeuroVascU-Net’s $\phi(\cdot)$7 and $\phi(\cdot)$8 fuse multi-scale context, edge cues, frequency-domain structure, and cross-domain feature streams within a single-modality 3D T1CE segmentation system (Liu et al., 4 Feb 2025, Vayeghan et al., 23 Nov 2025). These are relevant primarily as transferable fusion motifs rather than as direct multi-contrast modules.

6. Limitations, misconceptions, and current research directions

A common misconception is that any multi-input module qualifies as an MCFM. The surveyed literature does not support such a broad equivalence. Some modules fuse genuine image contrasts; others fuse contrast-adjusted versions of one image, temporal phases, or heterogeneous feature domains. In that sense, “multi-contrast” is task-specific rather than universal (Zhu et al., 13 Aug 2025, Li et al., 5 Dec 2025, Huang et al., 2024).

Another misconception is that more fusion is necessarily better. Recent multi-contrast MR reconstruction work identifies the opposite risk: target reconstruction can be contaminated by irrelevant reference information. MambaMDN is explicit on this point and therefore adopts a dual-domain strategy that first uses fully sampled reference $\phi(\cdot)$9-space to complete the undersampled target data and then applies a Mamba-based modality disentanglement network to extract and remove reference-specific features from the mixed representation, followed by iterative refinement (Lyu et al., 22 Dec 2025). This is a counterexample to the assumption that straightforward feature aggregation is always desirable.

A related limitation is alignment. ECFNet argues that direct cross-scale fusion is harmful unless scale mismatch is handled, because spatial and channel misalignment across layers introduces noise (Yang et al., 2024). FASR-Net similarly attributes degraded multi-contrast super-resolution to inappropriate foreground scale and patch size, motivating explicit flexible alignment before progressive fusion (Liu et al., 2022). In sequential fusion, registration across temporal phases remains critical; LIDIA registers all phases to the venous phase using DEEDS before feature fusion (Huang et al., 2024).

Generalization across acquisition conditions is another active issue. UniFS is motivated by the observation that existing multi-contrast MR reconstruction methods often require a separate model for each $$z^{(k)} = \frac{1}{C \times H' \times W'} \sum_{p=1}^{C} \sum_{i=1}^{H'} \sum_{j=1}^{W'} F^{(k)}_{pij},$$0-space undersampling pattern and may neglect frequency characteristics or extract only shallow frequency features. Its unified frequency-spatial fusion architecture is proposed specifically to handle multiple undersampling patterns, including previously unseen patterns, without retraining (Li et al., 5 Dec 2025). This suggests that current MCFM research is increasingly concerned not only with fusion efficacy but also with invariance to acquisition variability.

Taken together, the literature suggests that MCFM is best viewed as a design space rather than a single module archetype. Within that space, recurrent themes are early shallow fusion, selective branch weighting, multi-stage integration, cross-scale alignment, frequency-spatial decomposition, temporal iterative fusion, and modality-preserving refinement. The central research problem is no longer merely how to combine contrasts, but how to combine them in a way that is anatomically aligned, target-faithful, acquisition-robust, and resistant to redundant or reference-specific information transfer.