USRFNet: Ambiguous Neural Architecture Trends

• USRFNet is an umbrella term for distinct neural network models used in ultrasound imaging, microservice latency prediction, federated learning, and localization microscopy.
• In ultrasound beamforming, it replaces conventional delay-and-sum and Hilbert transform steps with a convolutional network that improves image quality and resolution.
• USRFNet also encompasses dual-stream and unrolled GNN-based optimizers, highlighting a context-dependent ambiguity that requires careful bibliographic disambiguation.

USRFNet is an overloaded research term used for several unrelated neural architectures in distinct technical domains. In ultrasound imaging, it denotes the “universal deep beamformer” introduced for variable-rate focused B-mode beamforming, a convolutional network that maps time-reversed, possibly sub-sampled RF channel data to complex beamformed I/Q outputs (Khan et al., 2019). In cloud systems, it denotes the “Unified System Representation Fusion Network,” a dual-stream model for predicting window-level P95 end-to-end latency in microservice systems by separately encoding traffic-side and resource-side signals before fusion (Qian et al., 3 Aug 2025). In federated learning, the term has been used for the unrolled GNN-based optimizer instantiated in the “Stochastic UnRolled Federated learning” framework, where the paper itself names the architecture U-DGD (Hadou et al., 2023). In ultrasound localization microscopy, the label does not formally appear as a method name, but may be used informally to describe an RF-trained super-resolution network that bypasses Delay-And-Sum beamforming (Hahne et al., 2023). The term therefore has no single canonical meaning; its interpretation depends entirely on disciplinary context.

1. Terminological scope and ambiguity

The literature assigns “USRFNet” to multiple architectures that share neither a task definition nor a mathematical formulation. The principal usages are summarized below.

Usage of “USRFNet”	Domain	Core role
Universal deep beamformer / DeepBF	Focused B-mode ultrasound	Maps delayed RF data to analytic beamformed I/Q data
Unified System Representation Fusion Network	Microservice performance modeling	Predicts window-level P95 end-to-end latency
U-DGD within SURF	Federated learning	Acts as a learned distributed optimizer
Informal “Ultrasound RF network” usage	Ultrasound localization microscopy	RF-domain super-resolution localization network

This multiplicity is not merely terminological. The ultrasound beamforming model replaces the summation-and-Hilbert-transform core of a focused B-mode pipeline (Khan et al., 2019). The microservice model learns a unified system embedding from a static service dependency graph, traffic features, and resource features (Qian et al., 3 Aug 2025). The federated learning model is not a task predictor at all, but an optimizer network unrolled from Distributed Gradient Descent (Hadou et al., 2023). In the ULM setting, the authors explicitly state that they “do not introduce a new acronym” and describe their method instead as an RF-trained super-resolution network based on mSPCN (Hahne et al., 2023).

A common misconception is that USRFNet denotes a single architecture reused across application areas. The published record indicates the opposite: the same label has been attached to conceptually independent systems. A plausible implication is that bibliographic disambiguation should rely on the associated paper title or arXiv identifier rather than on the acronym alone.

2. USRFNet as a universal deep beamformer in ultrasound imaging

In “Universal Deep Beamformer for Variable Rate Ultrasound Imaging” (Khan et al., 2019), USRFNet is a CNN that directly replaces the core of a focused B-mode ultrasound beamforming pipeline. Its inputs are time-reversed RF data from multiple receive channels and transmit events, for three adjacent depth samples, with arbitrary channel sub-sampling patterns; its outputs are the complex analytic beamformed signal for all scan lines at those depth samples. “Universal” means that the same network weights are used for different receive channel counts, different subsampling schemes, different specific random subsampling masks, and fully sampled data.

The model is positioned against two conventional families. Delay-and-sum beamforming is a fixed linear operation,

$$z_l[n] = \frac{1}{J}\mathbf{1}^\top \mathbf{y}_l[n],$$

or, with apodization,

$z_l[n] = \mathbf{w}_l[n]^\top \mathbf{y}_l[n].$

After beamforming, a Hilbert transform along depth gives the analytic signal $z_l^a[n]$, whose magnitude is used for envelope detection, followed by log compression. Minimum variance beamforming instead computes data-dependent weights from a covariance model. The paper emphasizes that such adaptive beamformers are computationally expensive, assume particular signal/noise statistics, and are difficult to adapt to arbitrary sparse or dynamically varying apertures (Khan et al., 2019).

USRFNet retains the physics-based time-reversal and delay calculation, but replaces both the summation stage and analytic signal construction with a learned nonlinear mapping. For each depth index $n$, the input is a $3\times64\times96$ Depth×Rx×TE tensor built from three adjacent depth planes in the Depth–Rx–TE subspace. The target output is represented as a $2\times3\times96$ tensor containing real and imaginary parts of the analytic beamformed signal across depth planes and scan lines. The CNN has 29 convolutional layers total; the first 28 layers use $3\times3$ convolutions followed by batch normalization and ReLU, and the final layer is a $1\times1$ convolution that maps features to the two output channels (Khan et al., 2019).

Formally, with parameters $\Theta$, the network approximates

$$\mathbf{z}^{a(i)}[n] = f_\Theta\big(\mathbf{Y}^{(i)}[n]\big),$$

and is trained by minimizing

$z_l[n] = \mathbf{w}_l[n]^\top \mathbf{y}_l[n].$0

The same weights are applied at all depths; depth is not explicitly encoded, although three adjacent depths are provided as context. This depth-independent formulation is later tied to one of the reported limitations, namely that axial resolution remains similar to DAS while lateral resolution improves (Khan et al., 2019).

3. Data, training, and reported behavior of the ultrasound beamformer

The beamforming USRFNet was trained on focused B-mode acquisitions from an Alpinion E-CUBE 12R scanner using a linear array L3-12H probe with center frequency 8.48 MHz, sampling frequency 40 MHz, 192 probe elements, 64 active receive elements per acquisition, 128 transmit elements, 96 transmit events, element pitch 0.2 mm, width 0.14 mm, and elevation length 4.5 mm (Khan et al., 2019). The in vivo carotid dataset comprised 10 volunteers with 40 frames per subject, for 400 temporal frames. From 4 subjects, 30,000 Rx–TE planes were randomly selected and assembled into $z_l[n] = \mathbf{w}_l[n]^\top \mathbf{y}_l[n].$1 cubes; the train/validation split was 25,000/5,000 samples. The remaining 360 frames were used as test data. An ATS-539 multipurpose tissue-mimicking phantom provided 188 additional frames used only for testing.

Training used variable sampling across depth only. For each training sample, the authors started from full 64-channel receive data and generated subsampled RF inputs with effective numbers of channels 64, 32, 24, 16, 8, and 4. At each depth plane, a random subset of channels was chosen under the given downsampling ratio, while always including the two channels at the center of the active transmit aperture. Targets were always computed from full-rate 64-channel DAS beamforming plus Hilbert transform, even when the inputs were subsampled. Optimization used MatConvNet in MATLAB 2015b, stochastic gradient descent, Xavier/Glorot Gaussian initialization, $z_l[n] = \mathbf{w}_l[n]^\top \mathbf{y}_l[n].$2 weight decay with parameter $z_l[n] = \mathbf{w}_l[n]^\top \mathbf{y}_l[n].$3, a learning rate decayed from $z_l[n] = \mathbf{w}_l[n]^\top \mathbf{y}_l[n].$4 to $z_l[n] = \mathbf{w}_l[n]^\top \mathbf{y}_l[n].$5, and 200 epochs (Khan et al., 2019).

Quantitatively, the paper reports consistent gains over DAS across in vivo carotid and phantom evaluations. For in vivo variable sampling, at 2× subsampling the CNR improved from 1.33 to 1.47, GCNR from 0.63 to 0.66, PSNR from 24.59 dB to 27.38 dB, and SSIM from 0.89 to 0.95; at 4× subsampling, CNR improved from 0.25 to 1.38, GCNR from 0.60 to 0.64, PSNR from 21.68 dB to 23.55 dB, and SSIM from 0.81 to 0.87 (Khan et al., 2019). The network was trained only on variable masks across depth, but similar relative improvements were reported for fixed masks not seen in training. The paper also states that the network can be applied to fully sampled RF data and improves image quality even at full rate.

On the phantom data, for variable sampling the average CNR values for DeepBF were 2.66, 2.65, 2.62, 2.53, 2.25, and 1.92 for 64, 32, 24, 16, 8, and 4 channels, respectively, and these were reported as 2.70%, 7.29%, 11.02%, 15.00%, 16.58%, and 14.29% higher than DAS. For fixed sampling, CNR improvements were 2.70%, 14.35%, 17.14%, 19.27%, 15.48%, and 13.16% over DAS across 64 to 4 channels. The authors also report graceful degradation of GCNR with subsampling and reconstruction time of approximately 9.8 ms per depth plane, which they describe as compatible with real-time imaging under parallelized or optimized implementations (Khan et al., 2019).

The paper frames these results as evidence that a single trained network can act as a learned nonlinear beamformer across multiple aperture sizes and subsampling patterns. This suggests that “universality” in this usage refers not to task generality, but to robustness across acquisition masks and sampling ratios within a fixed focused B-mode setting.

4. USRFNet as a unified system representation fusion network for microservices

In “Learning Unified System Representations for Microservice Tail Latency Prediction” (Qian et al., 3 Aug 2025), USRFNet denotes a dual-stream deep model for predicting window-level P95 end-to-end latency in microservice systems. The target is the 95th percentile latency of all user requests over a fixed 30-second window, with a 5-second sliding step. For each window $z_l[n] = \mathbf{w}_l[n]^\top \mathbf{y}_l[n].$6, the model learns

$z_l[n] = \mathbf{w}_l[n]^\top \mathbf{y}_l[n].$7

where $z_l[n] = \mathbf{w}_l[n]^\top \mathbf{y}_l[n].$8 is the static microservice dependency graph, $z_l[n] = \mathbf{w}_l[n]^\top \mathbf{y}_l[n].$9 and $z_l^a[n]$0 are traffic-side node and edge features, and $z_l^a[n]$1 are resource-side features.

The paper motivates this formulation by contrasting window-level P95 latency with per-request latency. It argues that per-request latency is volatile and operationally less useful, whereas window-level P95 is a smoothed, SLO-aligned signal that captures both system-wide trends and the “worst-case experience of the vast majority of users.” Two technical shortcomings in prior approaches are identified: inadequate handling of heterogeneous modalities and lack of principled designs for integrating them. Traffic-side features propagate across service dependencies and are volatile and event-driven; resource-side features are local to services or pods and evolve more slowly. Existing GNN methods are said to conflate these modalities when all metrics are processed in a single message-passing stream (Qian et al., 3 Aug 2025).

USRFNet addresses this by separating modeling and then fusing the resulting representations. The traffic-side encoder is a GNN with Transformer-based Graph Convolution layers, which uses the state graph $z_l^a[n]$2, node features $z_l^a[n]$3, and edge features $z_l^a[n]$4 to learn a traffic embedding $z_l^a[n]$5. The resource-side encoder is a gMLP over the resource matrix $z_l^a[n]$6, producing $z_l^a[n]$7. These are fused by the HIDAC module, “Hierarchical Integration of Demand And Capacity,” which first performs cross-diffusion attention,

$z_l^a[n]$8

and then low-rank multiplicative fusion,

$z_l^a[n]$9

followed by a final projection to the unified system embedding $n$0 (Qian et al., 3 Aug 2025).

The prediction head maps $n$1 to a scalar latency estimate. Training uses the Asymmetric Percentage Huber loss,

$n$2

with piecewise quadratic and linear regimes controlled by $n$3, $n$4, $n$5, and $n$6, and with $n$7 so that under-prediction is penalized more strongly than over-prediction. This asymmetry is justified as reflecting operational risk in latency underestimation (Qian et al., 3 Aug 2025).

5. Experimental profile of the microservice USRFNet

The microservice USRFNet was evaluated on two real-world microservice applications: Online Boutique with 11 microservices and Sockshop with 13 microservices. Both were deployed on a 3-node Kubernetes cluster with 1 master and 2 worker nodes, totaling 88 vCPUs and 256 GB RAM. Workloads were generated with Locust using gradual ramps, sharp spikes, sustained high load, and a probabilistic mixture of read-heavy and write-intensive user tasks. Metrics were collected through Istio and Prometheus, synchronized into 30-second windows with 5-second step, yielding 69,121 windows for Online Boutique and 71,696 windows for Sockshop. Data were split chronologically into 70% train, 10% validation, and 20% test (Qian et al., 3 Aug 2025).

The model’s input feature dimensions were $n$8 for traffic node features, $n$9 for traffic edge features, and $3\times64\times96$0 for resource features. The embedding dimension was $3\times64\times96$1 for both streams. The traffic-side GNN used 4 layers for Online Boutique and 3 for Sockshop; the resource-side gMLP used 4 blocks for Online Boutique and 5 for Sockshop; the HIDAC fusion rank was $3\times64\times96$2 for Online Boutique and $3\times64\times96$3 for Sockshop. Training used PyTorch 2.0.0, CUDA 11.8, a single NVIDIA RTX 4090, a 12 vCPU Xeon Platinum 8352V, 90 GB RAM, Adam with learning rate $3\times64\times96$4, batch size 32, dropout 0.1 in both encoders, and 500 epochs with early stopping (Qian et al., 3 Aug 2025).

On Online Boutique, the paper reports MAPE/MAE/RMSE of 9.85%/0.033 s/0.062 s for USRFNet, compared with 11.91%/0.042 s/0.084 s for GatedGCN$3\times64\times96$5, 12.19%/0.041 s/0.081 s for GCN$3\times64\times96$6, 12.37%/0.041 s/0.080 s for GIN$3\times64\times96$7, and 14.82%/0.056 s/0.102 s for GRAF. On Sockshop, USRFNet achieved 8.20%/0.029 s/0.071 s, compared with 9.17%/0.032 s/0.073 s for GatedGCN$3\times64\times96$8, 9.17%/0.033 s/0.075 s for GCN$3\times64\times96$9, 9.48%/0.034 s/0.076 s for GIN$2\times3\times96$0, and 9.75%/0.042 s/0.096 s for GRAF (Qian et al., 3 Aug 2025). The paper highlights relative MAPE reductions over the best GNN baseline of approximately 17.3% on Online Boutique and 10.6% on Sockshop.

The ablation study is central to the model’s interpretation. Traffic-only performance is reported as very poor; on Online Boutique, the example given is MAPE 32.12%. Resource-only performance is described as decent, with the example 10.75% MAPE on Online Boutique. Replacing the gMLP resource encoder with a GNN in the “GNN-fused” ablation degrades performance, and “Simple-fused,” which uses naive addition instead of HIDAC, is poorer than full USRFNet and in some metrics worse than Resource-only (Qian et al., 3 Aug 2025). The paper therefore treats specialization by modality and structured fusion as necessary design choices rather than incidental architectural details.

A plausible implication is that “unified system representation” in this context refers not to a generic multimodal embedding, but specifically to a representation in which demand propagation and capacity state are modeled with distinct inductive biases before interaction is learned.

6. USRFNet as an unrolled optimizer in federated learning

In “Stochastic Unrolled Federated Learning” (Hadou et al., 2023), the term USRFNet refers to the specific unrolled, GNN-based optimizer network instantiated in the SURF framework; the paper calls the architecture U-DGD. SURF is the meta-training method, whereas U-DGD is the unrolled architecture that implements a learned distributed optimizer for federated learning. The underlying optimization problem is

$2\times3\times96$1

over a connected graph $2\times3\times96$2, with the star graph as the special case corresponding to classical server-based federated learning.

The baseline algorithm being unrolled is Distributed Gradient Descent,

$2\times3\times96$3

SURF replaces fixed algorithmic updates with an unrolled neural network whose layers correspond to optimizer iterations. Each layer receives a stochastic mini-batch rather than the full local dataset, a design the paper terms stochastic unrolling. For decentralized FL, the U-DGD layer for agent $2\times3\times96$4 is

$2\times3\times96$5

where the graph filter

$2\times3\times96$6

acts as the GNN component and the second term is a shared local nonlinear update module (Hadou et al., 2023).

A defining feature of SURF is its layer-wise descent constraints. The meta-training problem minimizes expected downstream loss subject to

$2\times3\times96$7

The empirical dual problem is solved with a primal-dual algorithm. Under Assumptions 1–5, the paper proves near-optimality and near-feasibility guarantees through constrained learning theory, and derives an asymptotic result stating that the sequence of iterates visits a near-optimal region infinitely often, together with a finite-depth bound exhibiting exponential decay of expected gradient norms up to a residual term (Hadou et al., 2023).

Experimentally, the model used $2\times3\times96$8 unrolled layers and a graph filter order such as $2\times3\times96$9 in decentralized settings. The backbone task model was ResNet18 with convolutional layers pre-trained and frozen and a trainable softmax layer. Meta-training used 600 downstream CIFAR-10 datasets over 100 agents, with 45 training and 15 test examples per agent. In decentralized FL, USRFNet reportedly achieved higher accuracy in 20 communication rounds than DGD, DSGD, and DFedAvgM did in 200 rounds. In classical FL on a star graph with partial participation, it is reported to reach almost the same performance as centralized training in approximately 10 communication rounds, whereas other methods needed approximately 25 rounds to reach approximately 80% of centralized performance. An ablation without descent constraints showed near-zero accuracy until the final layer followed by a sudden jump, whereas constrained training improved test loss and accuracy gradually over layers (Hadou et al., 2023).

This usage differs categorically from the other USRFNet variants: the output is not an image or a latency estimate but an updated parameter state for distributed learning. The shared acronym therefore masks a fundamentally different ontological role for the network.

7. Informal RF-domain usage in ultrasound localization microscopy

In “Learning Super-Resolution Ultrasound Localization Microscopy from Radio-Frequency Data” (Hahne et al., 2023), there is no method explicitly named USRFNet. The authors describe their approach as “our RF-trained network” or “Ours (mSPCN+RF),” and state that they do not introduce a new acronym. If the term USRFNet is applied to this work, it is therefore an informal label for an RF-trained super-resolution network for ULM rather than an official model name.

The method uses the mSPCN architecture from Liu et al. (2020), but feeds raw complex RF channel data into the network instead of B-mode or IQ images. The input is

$3\times3$0

with real and imaginary parts stacked as channels, and the output is a super-resolved localization probability map

$3\times3$1

with upsampling factor $3\times3$2 in the main in silico experiments. Training minimizes

$3\times3$3

with $3\times3$4 and $3\times3$5 (Hahne et al., 2023).

A technical contribution of the paper is explicit B-mode↔RF coordinate transformation. Forward label projection maps a B-mode microbubble location $3\times3$6 to RF sample indices through a Time-of-Flight geometry,

$3\times3$7

and inverse point transformation from RF coordinates back to B-mode uses an approximate affine map estimated by Levenberg–Marquardt. This permits the network to produce a high-resolution localization map in RF coordinates while avoiding Delay-And-Sum beamforming entirely (Hahne et al., 2023).

On the PALA in silico benchmark, the RF-trained mSPCN achieved RMSE $3\times3$8 $3\times3$9, Jaccard 88.538%, and inference time 0.010 s, compared with RMSE $1\times1$0, Jaccard 93.748%, and 0.003 s + $1\times1$1 for mSPCN on B-mode/IQ, and RMSE $1\times1$2, Jaccard 87.883%, and 0.017 s + $1\times1$3 for U-Net on B-mode/IQ (Hahne et al., 2023). The authors interpret this as evidence that excluding DAS beamforming has potential to improve ULM localization accuracy. Because the acronym is not official in this paper, any encyclopedic use of “USRFNet” for this method should be treated as contextual shorthand rather than bibliographically canonical.

Across the literature, then, “USRFNet” names or informally denotes architectures that share only a broad family resemblance: each replaces a conventional hand-engineered stage with a learned representation or operator. In ultrasound beamforming, that stage is DAS plus Hilbert transform (Khan et al., 2019); in microservice performance modeling, it is monolithic feature processing in latency prediction pipelines (Qian et al., 3 Aug 2025); in federated learning, it is a fixed distributed optimizer (Hadou et al., 2023); and in RF-domain ULM, it is beamforming-centered localization (Hahne et al., 2023). This suggests that the term is best understood as a context-dependent acronym rather than as a stable model lineage.