USFM Framework: Multi-Domain Applications

• USFM Framework is a set of unified models that apply rigorous process mapping, dual-domain data analysis, and machine learning across manufacturing, medical imaging, and wireless communications.
• In manufacturing, USFM enables full traceability by linking sustainability goals to measurable KPIs via object process methodology, achieving clear energy and material balances.
• In medical imaging and wireless domains, USFM models demonstrate practical gains such as 2–8% improvement in Dice scores and a 3 dB BER advantage, underlining strong label efficiency and spectral performance.

The term "USFM Framework" encompasses several advanced frameworks across distinct technical domains, notably including: (1) the Unified Smart Factory Model for Industry 4.0 manufacturing systems, (2) Universal Ultrasound Foundation Models for label-efficient medical image analysis, and (3) Unified Sequency–Frequency Multiplexing for high-robustness wireless communication. Each instantiation is characterized by rigorous architecture, mathematical formalism, and empirical validation tailored for its respective field.

1. Unified Smart Factory Model (USFM) in Sustainable Manufacturing

The Unified Smart Factory Model (USFM) is a comprehensive, model-based framework designed to operationalize high-level sustainability goals—such as those described by the SDGs—into measurable, factory-level Key Performance Indicators (KPIs) via a systematic information architecture. USFM is articulated around three interlocking modules:

• Manufacturing Process and System: Models the entire physical and informational structure of the factory using Object Process Methodology (OPM). Each manufacturing, assembly, or auxiliary process is represented as an OPM process oval, with explicit inputs (raw materials, energy, information), outputs (products, waste), and enabling contextual links (equipment, human agents, environment). Objects in the system (e.g., material flows, equipment) possess measurable attributes (mass, energy flow, timestamps) essential for sustainability evaluation.
• Data Process: Implements a measurement-to-action pipeline consisting of Measure, Communicate, Store, Analyze, Decide, and Control sub-processes. These transform shop-floor data into actionable insights, supporting both semi-automated and automated shop-floor operations. The data process module receives its requirements directly from the Manufacturing Process module, ensuring that all collected shop-floor signals map onto relevant KPIs without redundancy.
• KPI Selection and Assessment: Facilitates a rigorous, ten-step procedure that links organizational or regulatory sustainability goals to quantifiable, factory-specific KPIs. This pipeline includes goal definition, KPI specification, metric formulation, data-source mapping, and decision feedback, ensuring traceability from high-level objectives down to sensor-level data collection.

OPM's representational rigor ensures completeness and transparency: missing sensors, logs, or data flows are readily identifiable. The model's information map links each process attribute to relevant KPIs and dictates sensor and data requirements at each process node.

Representative Material-Energy Balance

For any manufacturing process $P$ with inputs $\{I_k\}$ and outputs $\{O_l\}$, the material and energy balance is articulated as: $$\sum_k \dot m_{I_k} + \sum_k \dot E_{I_k} = \sum_l \dot m_{O_l} + \sum_l \dot E_{O_l} + W_P$$ where $\dot m$ are mass flow rates, $\dot E$ energy flow rates, and $W_P$ mechanical work (Kaushal et al., 11 Dec 2025).

KPI Mapping Procedure

The standard four-stage approach proceeds:

• Goal → KPI: (e.g., “Reduce carbon footprint” → CFU: carbon per unit produced)
• KPI → Metric:

$$\text{CFU} = \frac{1}{N}\left(\sum_{i\in \text{machines}} E_i\, EF_i + \sum_{j\in \text{materials}} m_j\, MF_j\right)$$

• Metric → Data Requirements: Defines which sensors (e.g., energy meters, weigh scales) and external data (e.g., emission factors from LCA databases) are required.
• Data Requirements → Factory Model: Each requirement is mapped to its corresponding OPM process/object node.

Case Study and Empirical Results

Applied to a PCB assembly SME, the USFM model enabled transparent measurement and optimization, with outcomes such as a Global Warming Potential of 89.21 kg CO$_2$e per kg PCB and identification of the reflow oven as the dominant energy consumer (≈90% line energy), focusing improvement efforts where most impactful (Kaushal et al., 11 Dec 2025).

2. Universal Ultrasound Foundation Model (USFM) in Medical Imaging

The Universal Ultrasound Foundation Model (USFM) is a large-scale, self-supervised vision transformer framework pre-trained to generalize across diverse ultrasound (US) imaging tasks and anatomical domains (Jiao et al., 2023, Gupta et al., 9 Jan 2026). The core architectural and methodological features include:

• Backbone: ViT-B/16, 12-layer transformer encoder ($D=768$), patch size $16\times 16$.
• Dual Masked Image Modeling (Spatial-Frequency MIM): During pre-training, both spatial mean-masking and frequency band-stop masking are employed. For spatial masking, random patches are replaced with the image mean, emulating speckle noise. For frequency masking, random annular bands are filtered in the DFT magnitude domain while preserving global structure: $$M_{\rm freq}(u,v) = \begin{cases} 0, & f_{k,\rm low} < \sqrt{u^2+v^2} < f_{k,\rm high} \ 1, & \text{otherwise} \end{cases}$$
• Self-Supervised Objective: The total loss comprises spatial $L_1$-reconstruction and focal frequency loss: $$\mathcal{L}_{\rm total} = \mathcal{L}_{\rm spa} + \lambda\,\mathcal{L}_{\rm freq}$$ with $\lambda=0.4$ determined by grid search.
• Organ-Balanced Sampling: Mini-batches are drawn according to organ-inverse-square-root frequency, ensuring rare organs are adequately represented.

Downstream Generalization and Label-Efficiency

• Plug-and-Play Adaptation: The frozen or fine-tuned foundation encoder readily supports segmentation (UPerNet), classification, and enhancement (CycleGAN with USFM branch).
• Performance: USFM outperforms ViT, CNN baselines, and SimMIM-pretrained on ImageNet by 2–8% in Dice and 2–7% in accuracy/F1 across 12 organs and several tasks. With only 20% labeled data, USFM matches full-data baseline performance, demonstrating strong label efficiency (Jiao et al., 2023).

Specialized Instantiations

• USFM in Tumor Segmentation: Employs a Vision Transformer backbone with attention-based multi-scale decoder (ATMHead), trained end-to-end with BCE+Dice hybrid loss (ATMLoss) for pancreatic tumor segmentation in EUS. Empirical Dice scores: $\sim0.66$, sensitivity $71.8\%$, specificity $97.7\%$ (Gupta et al., 9 Jan 2026).
• TinyUSFM: A lightweight distilled version of USFM using coreset selection (feature-gradient-driven K-means), domain-separated MIM, and per-head consistency-weighted knowledge distillation. Achieves $\geq 98\%$ of USFM benchmark accuracy on UniUS-Bench for both classification (84.91%) and segmentation (85.78%) using 6.36% of the parameters and 6.40% of the FLOPs (Ma et al., 22 Oct 2025).

3. USFM for Wireless Communication: Unified Sequency–Frequency Multiplexing

In the wireless communications domain, USFM stands for Unified Sequency–Frequency Multiplexing, a modulation scheme capitalizing on both frequency-domain (Fourier) and sequency-domain (Walsh-Hadamard) diversity (Alsulaimawi, 2024). The key technical components are:

• Joint Sequency–Frequency Transform (JSFT): Defines a two-dimensional transformation: $T = \mathrm{JSFT}\{G\} = F_N G W_N^T$ with $F_N$ as the unitary Fourier matrix and $W_N$ as the Walsh-Hadamard matrix. This enables symbol placement on a $(f, \sigma)$ grid for robust signaling.
• Machine-Learned CSI Adaptation: A small neural network is trained to map instantaneous Channel State Information (CSI) to optimal JSFT weighting, iteratively optimizing BER under stochastic gradient descent.

Theoretical and Empirical Results

• Error-Rate Superiority: USFM yields strictly lower average BER than OFDM in Rayleigh fading due to its joint-domain allocation, confirmed analytically and by simulation (3 dB BER gain at $10^{-3}$ versus OFDM).
• Spectral Efficiency: Achieves up to 18% higher spectral efficiency due to removal of the cyclic prefix and improved orthogonality.
• Processing Complexity: Overall per-frame complexity is $O(N^2\log N)$, with latency overhead mitigated via FPGA implementation (Alsulaimawi, 2024).
• Practical Guidelines: Includes recommendations for hardware (DSP/FPGA for fast FFT/WHT), grid sizing, and ML model size.

4. Data, Preprocessing, and Label Management

Each USFM variant emphasizes meticulous data collection, preprocessing, and sampling strategies:

• Industrial USFM: Requires detailed OPM diagrams for object, process, sensor, and metric annotation—enabling end-to-end traceability from sustainability objective to factory sensor.
• Medical USFM: Pretraining leverages datasets with 2.2M US images (classification, segmentation, enhancement), with on-the-fly augmentation (random rotation, crop, intensity jitter), plus organ-stratified batching.
• Wireless USFM: Simulation environments are parameterized over Rayleigh fading channels, Doppler shifts, and QAM symbol grids, while machine learning adaptation is trained on CSI-BER tuples.

5. Impact, Limitations, and Outlook

USFM frameworks have demonstrated notable advances in their respective sectors:

Domain	Primary Benefit	Representative Limitation
Manufacturing	Full traceability, data completeness, SME-friendliness	Implementation time, OPM adoption
Medical Imaging	Label-efficient, generalizes across organs, feature robustness	Still sensitive to out-of-distribution, domain shift for new US devices
Communications	Joint-domain robustness, higher spectral efficiency	Modest computational overhead (mitigable by hardware), requires ML inference for adaptation

Key limitations include the need for OPM familiarity and investment in process mapping for smart factories, the challenge of domain adaptation for new medical device types or imaging modalities, and, in wireless, additional inference for CSI-based adaptation.

A plausible implication is that methodologies pioneered in USFM (e.g., dual-domain masking, object/process modeling, ML-guided resource allocation) may be transferrable between disciplines, such as leveraging OPM-based information mapping in healthcare industrial automation or employing label-efficient foundation modeling for non-ultrasound modalities.

6. Key References

• Model-based manufacturing and sustainability: (Kaushal et al., 11 Dec 2025)
• Universal Ultrasound Foundation Model (USFM): (Jiao et al., 2023)
• USFM for scalable and efficient model distillation: (Ma et al., 22 Oct 2025)
• Ultrasound segmentation with USFM: (Gupta et al., 9 Jan 2026)
• Unified Sequency–Frequency Multiplexing in communications: (Alsulaimawi, 2024)
• Self-supervised ultrasound foundation model (USF-MAE): (Megahed et al., 27 Oct 2025)

These frameworks collectively position USFM as a flexible, technically rigorous approach adaptable to diverse domains, centered on the integration of domain-specific knowledge, efficient data and process modeling, and robust downstream performance.