RMFAT: Advanced Frameworks & Algorithms

• RMFAT is a set of advanced frameworks that apply specialized algorithms in communications, radio astronomy, cardiac MRI, video restoration, and radar detection.
• The methodologies employ robust statistical models, adaptive filtering, and spatio-temporal processing to enhance signal quality and system reliability.
• Each RMFAT instance integrates domain-specific techniques and benchmarking to address high-dimensional variability while delivering measurable performance gains.

RMFAT refers to several advanced frameworks and algorithms, each belonging to a distinct research area. Major utilizations include: (1) Rate-Splitting Multiple Access with Fluid Antenna Technology in communications, (2) Rotation Measure Flare Automatic Tracker for time-domain radio astronomy, (3) Respiratory/Motion-resolved FAT-fraction mapping in cardiac MRI, (4) Recurrent Multi-scale Feature Atmospheric Turbulence Mitigator for video restoration, and (5) Rapid Matched-Filter Adaptive Technique for radar detection. The following sections present comprehensive technical coverage for each RMFAT instance in its principal research context.

1. Rate-Splitting Multiple Access with Fluid Antenna Technology

System Model and Transmission Structure

In RMFAT for communications, the system comprises a base station (BS) with a single traditional antenna applying RSMA (Rate-Splitting Multiple Access) signaling to serve $K$ mobile users (MUs), each equipped with Fluid Antenna Systems (FAS) (Ghadi et al., 2024). Each message $w_k$ is split into a common part $w_c$ (decoded by all users) and a private part $w_{p,k}$ (decoded by user $k$ only). The two parts are linearly superposed: $$x = \sqrt{P}\Bigl(\sqrt{\alpha_c}s_c + \sum_{k=1}^K \sqrt{\alpha_{p,k}}s_{p,k}\Bigr)$$ with $\alpha_c + \sum_k\alpha_{p,k} = 1$ and $P$ the BS transmit power.

On the user side, each MU employs a 2D FAS of area $W_k\lambda^2$ with $N_k = N_k^1 \times N_k^2$ reconfigurable ports. In each channel use, the port maximizing instantaneous received power is activated: $w_k$0 where the channel is modeled as correlated Rayleigh fading, with spatial correlations described by Bessel functions.

After successively decoding the common and private streams, the relevant SINRs are: $w_k$1 with SNR $w_k$2 and path loss $w_k$3.

Channel Gain Statistics and Outage Analysis

The equivalent channel power gain $w_k$4 under FAS is statistically characterized via a multivariate Student-$w_k$5 copula: $w_k$6 where $w_k$7 and $w_k$8 encodes port correlation. The per-user outage probability, accounting for both common and private stream SINR, is given by: $w_k$9 where $w_c$0 is a function of SINR targets and system parameters.

At high SNR, the diversity order achieves $w_c$1 (number of FAS ports), with coding gain dependent on dependence structure and copula tails.

Performance Comparison and Conclusions

Numerical evaluation demonstrates that FAS-RSMA (RMFAT) achieves ultra-reliable outage ($w_c$2 at $w_c$3 dB), outperforming TAS-based RSMA and FAS-NOMA by orders of magnitude. The diversity benefit is fully realizable by increasing FAS port count or effective surface area. RMFAT delivers massive spatial diversity and flexible interference management, yielding dramatically lower outage (Ghadi et al., 2024).

2. Rotation Measure Flare Automatic Tracker in Fast Radio Burst Analysis

Pipeline Foundation and Modeling

RMFAT for FRB polarization exploits time-series analysis to detect statistically significant, localized deviations ("RM flares") in the observed rotation measure (RM) of repeating fast radio bursts (Yang et al., 16 Apr 2026). The observed series is decomposed as: $w_c$4 with $w_c$5 as quiescent background, $w_c$6 as transient flare, and $w_c$7 as heteroscedastic noise.

Foreground flares manifest as excursions beyond an adaptively estimated, robust rolling-median baseline. The noise model incorporates instrumental errors, local volatility, and a global MAD-derived floor.

Detection Algorithm

Significance scores are calculated: $w_c$8 with candidate flares defined by the maximum $w_c$9 exceeding a high threshold $w_{p,k}$0 (default 10). Flare intervals are defined using a full-width-at-tenth-maximum criterion, ensuring robust boundary assignment.

The method incorporates rigorous outlier rejection and adaptively sized filters, with false-alarm probabilities computed via standard Gaussian approximations and consideration of correlations.

Performance and Impact

Applied uniformly to 15 repeating FRB sources, only one source (FRB 20220529A) showed a high-confidence RM flare (peak $w_{p,k}$1RM $w_{p,k}$2 2000 rad m$w_{p,k}$3 over ~14 days). All other sources exhibited complex RM variations below the detection threshold, affirming the rarity of such events. The pipeline yields near-zero false positive rates at conservative thresholds and distinguishes genuine flares from secular or stochastic RM evolution (Yang et al., 16 Apr 2026).

3. Respiratory/Motion-resolved FAT-fraction Mapping in Cardiac MRI

Sequence and Acquisition Strategy

RMFAT in cardiac MRI refers to a free-running multi-echo 3D radial GRE protocol for simultaneous cardiac- and respiratory-motion-resolved, whole-heart fat-fraction mapping (Mackowiak et al., 2022). The framework integrates Pilot Tone (PT) navigation for robust gating and XD-GRASP compressed sensing for image reconstruction.

Key protocol details:

• 1.5 T system, $w_{p,k}$4 mm$w_{p,k}$5 isotropic, FOV $w_{p,k}$6 mm$w_{p,k}$7
• Multiecho (NTE $w_{p,k}$8) monopolar fly-back, TR $w_{p,k}$9 ms, TE$k$0 $k$1 ms, $k$2TE $k$3 ms
• Scan time $k$4 6 min 15 s, with each bin capturing $k$5 of full Nyquist.

Motion Extraction and Image Reconstruction

PT navigation captures cardiac and respiratory signals by analyzing a continuous-wave RF pilot at 62 MHz. Bins are assigned based on normalized, filtered PT time-series, yielding high correlation against self-gating and ECG.

Data are reconstructed as a 6D array using XD-GRASP regularized along motion axes: $k$6 No echo-dimension regularization is used.

Fat–Water Quantification

Each motion bin’s multi-echo signal is modeled for water and multi-peak fat using maximum likelihood and graph-cut optimization: $k$7 Fractional fat maps are computed as: $k$8

Results and Observations

Compared to ECG-triggered reference, motion-resolved RMFAT achieves:

• Bland-Altman bias of $k$9, $$x = \sqrt{P}\Bigl(\sqrt{\alpha_c}s_c + \sum_{k=1}^K \sqrt{\alpha_{p,k}}s_{p,k}\Bigr)$$0 correlation for pericardial fat fraction;
• Accurate tracking of physiological pericardial FF decrements of $$x = \sqrt{P}\Bigl(\sqrt{\alpha_c}s_c + \sum_{k=1}^K \sqrt{\alpha_{p,k}}s_{p,k}\Bigr)$$1 at systole;
• Significant FF underestimation for $$x = \sqrt{P}\Bigl(\sqrt{\alpha_c}s_c + \sum_{k=1}^K \sqrt{\alpha_{p,k}}s_{p,k}\Bigr)$$2 vs $$x = \sqrt{P}\Bigl(\sqrt{\alpha_c}s_c + \sum_{k=1}^K \sqrt{\alpha_{p,k}}s_{p,k}\Bigr)$$3, confirming importance of sufficiently sampled echoes.

PT-based navigation surpasses conventional self-gating in trigger loss rate and temporal jitter. The main limitation is the high computational workload, with total reconstruction time $$x = \sqrt{P}\Bigl(\sqrt{\alpha_c}s_c + \sum_{k=1}^K \sqrt{\alpha_{p,k}}s_{p,k}\Bigr)$$4 hours (Mackowiak et al., 2022).

4. Recurrent Multi-scale Feature Atmospheric Turbulence Mitigator

Model Design

In video restoration under atmospheric turbulence, RMFAT is a recurrent, multi-scale spatio-temporal framework that processes only the current distorted frame and the previously restored frame at each step (Liu et al., 15 Aug 2025). Inputs are concatenated as: $$x = \sqrt{P}\Bigl(\sqrt{\alpha_c}s_c + \sum_{k=1}^K \sqrt{\alpha_{p,k}}s_{p,k}\Bigr)$$5 Features are processed via a hierarchically stacked 3D convolution and Transformer encoder–decoder, with temporal warping and fusion at all scales.

Temporal Warping and Loss Design

Temporal alignment is performed recursively: $$x = \sqrt{P}\Bigl(\sqrt{\alpha_c}s_c + \sum_{k=1}^K \sqrt{\alpha_{p,k}}s_{p,k}\Bigr)$$6 Feature refinement employs residual attention blocks and layer normalization.

The loss function integrates spatial fidelity (Charbonnier loss), wavelet-based detail preservation, flow-based temporal consistency, and a detection-guided YOLO objective.

Benchmarking and Efficacy

RMFAT achieves:

• SSIM $$x = \sqrt{P}\Bigl(\sqrt{\alpha_c}s_c + \sum_{k=1}^K \sqrt{\alpha_{p,k}}s_{p,k}\Bigr)$$7 0.686 on synthetic-dynamic and $$x = \sqrt{P}\Bigl(\sqrt{\alpha_c}s_c + \sum_{k=1}^K \sqrt{\alpha_{p,k}}s_{p,k}\Bigr)$$8 on real-dynamic data,
• $$x = \sqrt{P}\Bigl(\sqrt{\alpha_c}s_c + \sum_{k=1}^K \sqrt{\alpha_{p,k}}s_{p,k}\Bigr)$$99% SSIM gain over the best prior art,
• Inference speed %%%%48$k$49%%%% faster than comparable architectures (0.008 s/frame, 2.6 M parameters).

Ablation studies indicate strong sensitivity of performance to temporal warping and recurrent architecture. Limiting factors include degraded correction for extremely rapid, transient objects and very strong turbulence (Liu et al., 15 Aug 2025).

5. Rapid Matched-Filter Adaptive Technique for Radar Target Detection

Doppler-Domain Localized Adaptive Matched Filter (DDL-AMF)

RMFAT in radar is formalized as a Doppler-domain localized adaptive matched filter (DDL-AMF), leveraging the concept of a Region of Possible Target Detection (RPTD) (Kononov, 20 May 2026).

Instead of full-spectrum covariance adaptation, RMFAT confines filter adaptation to a small set of $\alpha_c + \sum_k\alpha_{p,k} = 1$2 Doppler bins that capture most target power: $\alpha_c + \sum_k\alpha_{p,k} = 1$3 where $\alpha_c + \sum_k\alpha_{p,k} = 1$4 is the DFT of the steering vector in these bins, $\alpha_c + \sum_k\alpha_{p,k} = 1$5 the observed signal, and $\alpha_c + \sum_k\alpha_{p,k} = 1$6 the local covariance estimated from $\alpha_c + \sum_k\alpha_{p,k} = 1$7 secondary cells.

Detection Process and Computational Efficiency

The RPTD selection algorithm maximizes energy capture around a hypothesized Doppler frequency, minimizing window width. RMFAT then applies the AMF on this localized region, requiring no knowledge of clutter spectrum or RODI parameterization.

Complexity is reduced from $\alpha_c + \sum_k\alpha_{p,k} = 1$8 (full AMF) to $\alpha_c + \sum_k\alpha_{p,k} = 1$9 per CUT, achieving $P$0–$P$1 speedup. Near-optimum detection probability ($P$2 below ideal) is retained even in multimodal clutter, unknown Doppler, or under severe data limitations.

Theoretical Guarantees and Performance

False-alarm control remains CFAR via a threshold linked to the Beta distribution for small $P$3:

$P$4

Empirical and analytic results confirm robust detection and ROC within 1 dB of optimal full-dimension AMF (Kononov, 20 May 2026).

---

Each RMFAT variant constitutes a distinct, rigorously tested framework tailored to its domain, with common emphasis on robust statistical modeling, efficient inference, and practical applicability in the presence of high-dimensionality, variability, or resource constraints.