ClinicalFMamba: Multimodal Neuroimaging Fusion

• ClinicalFMamba is a clinically oriented Mamba-based modeling paradigm that fuses multimodal neuroimaging data using selective state-space techniques.
• It combines CNNs, language-grounded reasoning, and efficient scanning patterns to capture long-range dependencies while preserving local anatomical details.
• The architecture supports diverse clinical tasks such as brain tumor classification, deformable registration, and longitudinal prediction with proven performance metrics.

ClinicalFMamba denotes a clinically oriented line of Mamba-based modeling built around selective state-space models for medical imaging and adjacent clinical reasoning tasks. In current usage, the term has a dual meaning. It names a specific end-to-end CNN–Mamba hybrid for multimodal neuroimaging fusion, and it also denotes a broader paradigm in which linear-time selective scanning is combined with domain-specific mechanisms such as frequency-domain encoding, graph priors, adaptive conditioning, weak supervision, multimodal fusion, or language-grounded reasoning. Across these formulations, the central aim is to preserve long-range dependency modeling without the quadratic cost of Transformer self-attention while maintaining the local detail, anatomical fidelity, and computational practicality demanded by clinical workflows (Zhou et al., 5 Aug 2025, Cheng et al., 9 Jun 2025, Yang et al., 15 May 2025, Guo et al., 2024).

1. Terminological scope and representative domains

In the narrow sense, ClinicalFMamba is the model introduced in "ClinicalFMamba: Advancing Clinical Assessment using Mamba-based Multimodal Neuroimaging Fusion," which targets 2D and 3D multimodal fusion and downstream brain tumor classification. In the broader sense used across related papers, ClinicalFMamba describes clinically oriented Mamba systems for weakly supervised histopathology, deformable registration, disease trajectory prediction, ophthalmic angiography classification, multi-view radiography, medical segmentation, and neuroimaging-plus-LLM reasoning (Zhou et al., 5 Aug 2025, Cheng et al., 9 Jun 2025, Guo et al., 2024, Yang et al., 15 May 2025, Wang et al., 28 Jan 2026, Zheng et al., 4 Mar 2025, Xu, 2024, Yang et al., 24 May 2025, Rong et al., 26 Jul 2025, Torabi et al., 14 Feb 2026, Fang et al., 2024).

System	Clinical task	Defining additions
FMaMIL (Cheng et al., 9 Jun 2025)	Weakly supervised histopathology lesion segmentation	MIL, LFDE, CAM-guided two-stage refinement
MambaMorph (Guo et al., 2024)	MR–CT and T1–T2 deformable registration	Fine-grained feature extractor, Mamba registration, STN
MambaControl (Yang et al., 15 May 2025)	Diffusion-based AD trajectory prediction	Anatomy graph control, Fourier spectral refinement
CLEAR-Mamba (Wang et al., 28 Jan 2026)	Ophthalmic angiography classification	HaC, RaP, MedMamba/VSSM backbone
XFMamba (Zheng et al., 4 Mar 2025)	Multi-view medical image classification	CVSM and MVCM cross-fusion
ClinicalFMamba (Zhou et al., 5 Aug 2025)	Multimodal neuroimaging fusion	DGCB, tri-plane scanning, CMCA
HC-Mamba (Xu, 2024)	Medical image segmentation	Hybrid convolutions with SS2D
TK-Mamba (Yang et al., 24 May 2025)	Text-driven 3D segmentation	EGSC, ToM, 3D-GR-KAN, CLIP branches
FaRMamba (Rong et al., 26 Jul 2025)	Frequency-aware segmentation	MSFM and SSRAE
NeuroMambaLLM (Torabi et al., 14 Feb 2026)	Dynamic fMRI analysis and report generation	Dynamic latent graph learning, LoRA-aligned LLM
GFE-Mamba (Fang et al., 2024)	MCI-to-AD progression prediction	MRI→PET GFE, BCA, multimodal Mamba classifier

This usage pattern indicates that ClinicalFMamba is not restricted to a single modality or output type. It spans fusion, segmentation, classification, registration, longitudinal prediction, and explanatory reporting, provided that the system treats Mamba-style selective state-space modeling as the primary mechanism for scalable long-range dependency capture.

2. State-space foundations and recurrent design patterns

The shared mathematical substrate is the state-space model. Several papers express the continuous-time form as

$\dot{x}(t)=Ax(t)+Bu(t), \qquad y(t)=Cx(t)+Du(t),$

with zero-order-hold discretization yielding a recurrence of the form

$$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$

Mamba makes this formulation selective by allowing parameters such as $\Delta$, $B$, and $C$ to depend on the current token or feature, thereby producing input-dependent gating and linear-time sequence mixing rather than quadratic attention (Zhou et al., 5 Aug 2025, Cheng et al., 9 Jun 2025, Guo et al., 2024, Zheng et al., 4 Mar 2025, Yang et al., 15 May 2025).

In vision and clinical imaging, this general mechanism is specialized through different scan topologies. SS2D and related visual state-space modules scan 2D feature maps along multiple directions; FMaMIL uses row-wise and column-wise bidirectional scanning for bags of histopathology patches; ClinicalFMamba extends scanning to axial, coronal, and sagittal planes for 3D volumetric fusion; XFMamba applies SS2D within multi-view encoders; MambaMorph serializes volumetric patches into long sequences for deformable registration (Cheng et al., 9 Jun 2025, Zhou et al., 5 Aug 2025, Zheng et al., 4 Mar 2025, Guo et al., 2024).

A second recurring pattern is the explicit reintroduction of locality after serialization. Clinical imaging papers repeatedly note that tokenization and 1D scanning can erode local adjacency or high-frequency detail. The compensatory mechanisms differ by task: DGCB and CMCA in ClinicalFMamba, dilated and depthwise separable convolutions in HC-Mamba, EGSC and 3D-GR-KAN in TK-Mamba, LFDE in FMaMIL, and MSFM with SSRAE in FaRMamba (Zhou et al., 5 Aug 2025, Xu, 2024, Yang et al., 24 May 2025, Cheng et al., 9 Jun 2025, Rong et al., 26 Jul 2025). This suggests that ClinicalFMamba is best understood as a design pattern in which Mamba supplies global sequence mixing, while auxiliary modules restore spatial fidelity, semantic structure, or modality alignment.

3. Task-specific formulations

In weakly supervised histopathology, FMaMIL formulates a whole-slide image or cropped clinical image as a bag of non-overlapping patches encoded by a pre-trained CNN and processed by a Mamba-based MIL encoder. Its learnable frequency-domain encoding operates along the token axis using 1D FFT, learnable complex weights, inverse FFT, normalization, and gating, after which CAMs are thresholded at $t=0.5$ and refined by a U-Net trained with soft-label supervision and confidence-weighted self-correction. On the private glomerular dataset, the reported validation results are classification Acc$=0.996$, AUC$=0.998$, and segmentation mIoU$=0.887$, Dice$=0.934$; on CAMELYON16, classification Acc$$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$0, AUC$$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$1, and segmentation mIoU$$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$2, Dice$$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$3 (Cheng et al., 9 Jun 2025).

In deformable registration, MambaMorph separates feature extraction from registration. A simple U-Net-like CNN with one down-sampling step and fixed 16-channel layers extracts modality-specific features, which are concatenated and passed to a Mamba-based registration module that predicts a dense 3D deformation field and uses an STN for warping. On SR-Reg, the reported test Dice is $$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$4 with HD95 $$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$5 mm, and on IXI the reported Dice is $$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$6 with HD95 $$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$7 (Guo et al., 2024).

In longitudinal prediction, MambaControl combines a Mamba diffusion backbone with a ControlNet-like pathway that computes graph-based anatomical features refined in the spectral domain. The graph is built from patch features, converted into a normalized Laplacian, and filtered by a Chebyshev-approximated spectral operator $$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$8. On ADNI-3, the Fourier graph variant reports SSIM $$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$9, PSNR $\Delta$0 dB, and regional MAEs of $\Delta$1 in hippocampus, $\Delta$2 in amygdala, $\Delta$3 in lateral ventricle, $\Delta$4 in thalamus, and $\Delta$5 in CSF (Yang et al., 15 May 2025).

In classification, CLEAR-Mamba keeps the MedMamba backbone but adds HaC, a hypernetwork-based adaptive conditioning layer, and RaP, an evidential Dirichlet head. The evidential formulation uses

$\Delta$6

with a loss combining marginal likelihood and a KL prior regularizer to $\Delta$7. On the in-house FFA/ICGA dataset, CLEAR-B reports OA $\Delta$8, F1 $\Delta$9, Specificity $B$0, and AUC $B$1; on OCT-C8, CLEAR-S reports OA $B$2 and AUC $B$3 (Wang et al., 28 Jan 2026). XFMamba addresses unregistered multi-view classification through channel interleaving and shared-decoder fusion, reporting AUROC $B$4 on MURA for XFMamba-S, $B$5 on CheXpert for XFMamba-B, and $B$6 on CBIS-DDSM for XFMamba-B (Zheng et al., 4 Mar 2025).

In segmentation, HC-Mamba combines SS2D with a hybrid-convolution branch, TK-Mamba extends Mamba to text-driven 3D segmentation with ToM and 3D-GR-KAN, and FaRMamba explicitly repairs high-frequency loss and 2D structure degradation with MSFM and SSRAE. These formulations differ in detail but share the same principle: Mamba handles long-range context, while explicit local or semantic modules repair structures that serialization alone can weaken (Xu, 2024, Yang et al., 24 May 2025, Rong et al., 26 Jul 2025).

4. ClinicalFMamba as multimodal neuroimaging fusion architecture

The specific system named ClinicalFMamba targets multimodal image fusion in MRI–CT, MRI–SPECT, and 2D/3D MRI fusion on BraTS 2019, followed by downstream LGG-versus-HGG classification. Its pipeline contains two modality-specific hybrid feature encoders built around Dilated Gated Convolution Blocks, a latent Mamba fusion stage, Cross-Modal Channel Attention in decoding, and a lightweight convolutional decoder. In 2D, the latent stage uses four-directional Mamba scanning plus Fusion Mamba blocks; in 3D, it introduces tri-plane selective scanning over axial, coronal, and sagittal slice sequences, followed by Fusion Mamba aggregation and CMCA-enhanced decoding (Zhou et al., 5 Aug 2025).

The DGCB uses parallel $B$7 and $B$8 convolutions, a gated interaction $B$9, and dilated convolutions with rates $C$0 and $C$1. The fusion loss is unsupervised and combines pixel, gradient, and SSIM terms:

$C$2

with $C$3, $C$4, and $C$5. The downstream classifier is a 2D or 3D ResNet-50 trained with focal loss rather than joint multitask optimization.

Quantitatively, the reported 2D MRI–CT fusion results are PSNR $C$6, SSIM $C$7, FMI $C$8, FSIM $C$9, and EN $t=0.5$0; the MRI–SPECT results are PSNR $t=0.5$1, SSIM $t=0.5$2, FMI $t=0.5$3, FSIM $t=0.5$4, and EN $t=0.5$5. In 3D T2–FLAIR fusion on BraTS, ClinicalFMamba-3D reports PSNR $t=0.5$6, MS-SSIM $t=0.5$7, and EN $t=0.5$8. Downstream classification on fused BraTS data yields AUC $t=0.5$9, F1 $=0.996$0, and Acc $=0.996$1 in 2D, and AUC $=0.996$2, F1 $=0.996$3, and Acc $=0.996$4 in 3D. The 2D model has about $=0.996$5M parameters and runtime about $=0.996$6 s per image pair at $=0.996$7; the 3D variant has about $=0.996$8M parameters and runtime about $=0.996$9 s per $=0.998$0 volume pair. Ablations report that removing CMCA degrades MRI–CT fusion by $=0.998$1 PSNR, $=0.998$2 SSIM, $=0.998$3 FMI, $=0.998$4 FSIM, and $=0.998$5 EN, while replacing tri-plane scanning with 2D scanning on BraTS-3D degrades PSNR by $=0.998$6, MS-SSIM by $=0.998$7, and EN by $=0.998$8 (Zhou et al., 5 Aug 2025).

5. Quantitative performance and efficiency profile

A notable feature of the ClinicalFMamba literature is that quantitative gains are reported across highly heterogeneous tasks rather than within a single benchmark family.

Model	Dataset-task	Reported result
FMaMIL (Cheng et al., 9 Jun 2025)	CAMELYON16 weakly supervised segmentation	mIoU $=0.998$9, Dice $=0.887$0
MambaMorph (Guo et al., 2024)	SR-Reg MR–CT registration	Dice $=0.887$1, HD95 $=0.887$2 mm
MambaControl (Yang et al., 15 May 2025)	ADNI trajectory prediction	SSIM $=0.887$3, PSNR $=0.887$4 dB
CLEAR-S (Wang et al., 28 Jan 2026)	OCT-C8 classification	OA $=0.887$5, AUC $=0.887$6
ClinicalFMamba-3D (Zhou et al., 5 Aug 2025)	BraTS T2–FLAIR fusion	PSNR $=0.887$7, MS-SSIM $=0.887$8
HC-Mamba (Xu, 2024)	Synapse multi-organ CT segmentation	Dice $=0.887$9, HD95 $=0.934$0
TK-Mamba (Yang et al., 24 May 2025)	MSD+KiTS23 average	Dice $=0.934$1, NSD $=0.934$2
FaRMamba-DCT (Rong et al., 26 Jul 2025)	Kvasir-Seg segmentation	DSC $=0.934$3, MIoU $=0.934$4
NeuroMambaLLM (Torabi et al., 14 Feb 2026)	ABIDE I ASD vs TC	Accuracy $=0.934$5, F1 $=0.934$6
GFE-Mamba (Fang et al., 2024)	ADNI 1-year MCI→AD prediction	Accuracy $=0.934$7, MCC $=0.934$8

Efficiency claims are similarly task-specific. MambaMorph reports runtime $=0.934$9 s per volume pair, memory about $$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$00 GB at test time, and about $$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$01 M parameters; ClinicalFMamba reports real-time 2D fusion and feasible 3D fusion with small parameter counts; CLEAR-Mamba spans $$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$02M, $$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$03M, and $$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$04M parameter scales; HC-Mamba reduces parameters to $$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$05M in the hybrid-convolution ablation; TK-Mamba trains on $$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$06 crops with batch size $$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$07 on a single RTX 4090; FaRMamba trains with batch size $$x_t=\bar{A}x_{t-1}+\bar{B}u_t, \qquad y_t=Cx_t+Du_t.$$08 on a single RTX 3090 (Guo et al., 2024, Zhou et al., 5 Aug 2025, Wang et al., 28 Jan 2026, Xu, 2024, Yang et al., 24 May 2025, Rong et al., 26 Jul 2025). The common empirical claim is therefore not merely higher accuracy, but higher accuracy under practical memory or latency constraints.

6. Limitations, misconceptions, and future directions

A frequent misconception is that ClinicalFMamba names a single standardized architecture. The literature does not support that interpretation. Some systems are explicitly frequency-driven, such as FMaMIL, MambaControl, and FaRMamba; others prioritize adaptive conditioning and evidential uncertainty, as in CLEAR-Mamba; others focus on missing-modality compensation or language-grounded reporting, as in GFE-Mamba and NeuroMambaLLM (Cheng et al., 9 Jun 2025, Yang et al., 15 May 2025, Rong et al., 26 Jul 2025, Wang et al., 28 Jan 2026, Fang et al., 2024, Torabi et al., 14 Feb 2026).

A second misconception is that linear-time selective scanning automatically resolves clinical trustworthiness. The papers themselves document residual limitations. MambaMorph predicts displacement fields directly and uses no explicit stationary velocity field or Jacobian penalty, with non-positive Jacobian percentages remaining small but nonzero. CLEAR-Mamba emphasizes calibrated evidential outputs but does not report ECE, NLL, or Brier in the tables. The ClinicalFMamba fusion paper reports no reader studies and does not model uncertainty. NeuroMambaLLM treats generated reports as exploratory analysis rather than formal diagnosis (Guo et al., 2024, Wang et al., 28 Jan 2026, Zhou et al., 5 Aug 2025, Torabi et al., 14 Feb 2026).

Several limitations recur across task families. FMaMIL depends on bag construction, patch ordering, and imperfect CAMs in extremely ambiguous lesions, while its current LFDE operates along the token axis rather than explicitly multi-scale 2D spectra per patch. MambaControl is sensitive to patching scheme, adjacency definition, Laplacian normalization, and spectral filter order, and may require recalibration under domain shift. TK-Mamba notes that CLIP text embeddings come from general web corpora and may limit semantic precision for rare medical conditions; its multi-task setting can also underperform on some kidney structures relative to SegMamba. FaRMamba requires modality-appropriate transform selection, since DWT, FFT, and DCT are reported to behave differently on ultrasound, MRI, and endoscopy (Cheng et al., 9 Jun 2025, Yang et al., 15 May 2025, Yang et al., 24 May 2025, Rong et al., 26 Jul 2025).

The future directions stated across the literature are correspondingly diverse. They include multi-scale spectral fusion and pathology priors for weakly supervised pathology, diffeomorphic constraints and Jacobian regularization for registration, multi-center prospective validation and explicit risk-coverage analysis for ophthalmic deployment, extension to PET, DWI, perfusion, and quantitative maps for neuroimaging fusion, 3D transforms and multi-axis scanning for frequency-aware segmentation, and broader disease transfer for dynamic graph plus LLM systems (Cheng et al., 9 Jun 2025, Guo et al., 2024, Wang et al., 28 Jan 2026, Zhou et al., 5 Aug 2025, Rong et al., 26 Jul 2025, Torabi et al., 14 Feb 2026). A plausible implication is that ClinicalFMamba is evolving less toward a monolithic foundation model than toward a modular clinical systems framework: selective state-space modeling forms the computational core, while clinically specific priors determine the final architecture.