BEAM-Net: A Context-Sensitive Neural Framework

• BEAM-Net is a context-sensitive label for deep-learning neural networks applied to beamforming across wireless communications, ultrasound imaging, and heavy-ion fluctuation studies.
• In wireless communications, BEAM-Net uses both data-driven and model-driven approaches to replace high-complexity beamforming with fast, near-optimal neural predictions under QoS and power constraints.
• In ultrasound imaging, BEAM-Net employs an end-to-end U-Net architecture with an integrated Bone Probability Map to enhance image quality, contrast, and structural details.

Searching arXiv for the cited BEAM-Net-related papers to ground the article. arxiv_search(query="(Xia et al., 2020)", max_results=5) BEAM-Net is a non-unique term in recent technical literature. In wireless communications, it is used as an umbrella term for deep-learning-based beamforming solutions, called beamforming neural networks (BNNs), for multiuser MIMO downlink. In ultrasound imaging, it is the formal name of an end-to-end deep neural network for high-resolution, high-frame-rate musculoskeletal ultrasound beamforming with integrated bone enhancement. In a separate heavy-ion context, it is used as a shorthand for Beam Energy dependence of Moments of net-charge results. By contrast, the optical mode-demultiplexing method BPNet is explicitly not termed BEAM-Net (Xia et al., 2020, Madhusoodanan et al., 21 Jul 2025, Collaboration, 2014, Bekerman et al., 2019).

1. Terminological scope and reuse

In the 2020 wireless paper, BEAM-Net and BNN refer to the same concept introduced therein: an NN-based predictor of beamforming or precoding for multiuser MIMO downlink. The central motivation is to replace high-complexity, iterative beamforming optimization—whose latency makes it ill-suited for fast-fading wireless channels—with a trained neural network that performs near-optimal beamforming via fast inference. The paper distinguishes data-driven and model-driven BEAM-Net, with the latter embedding explicit signal-processing modules built from expert knowledge such as uplink–downlink duality, normalization, and conversion steps (Xia et al., 2020).

In the 2025 ultrasound paper, BEAM-Net is a novel end-to-end deep neural network that performs high-frame-rate ultrasound beamforming with integrated bone enhancement, using single-plane-wave radio frequency data as input. Its defining feature is a Bone Probability Map that acts as an attention mechanism to enforce higher structural similarity around bony regions in the image. The paper describes this as the first of its kind to incorporate bone enhancement directly into ultrasound beamforming using deep learning (Madhusoodanan et al., 21 Jul 2025).

The same label is therefore domain-dependent. A plausible implication is that citation context is necessary whenever BEAM-Net is used without qualification, because the term can denote distinct frameworks in wireless communications and medical ultrasound, and can also appear as shorthand in a heavy-ion fluctuation study.

2. BEAM-Net in wireless communications

In the wireless literature, BEAM-Net denotes a beamforming neural network for downlink MU-MIMO, with the paper’s examples focusing on MISO downlink. The general system model uses $K$ single-antenna users, $M$ transmit antennas at the base station, channel matrix $H \in \mathbb{C}^{K \times M}$, precoder $W \in \mathbb{C}^{M \times K}$, and user symbols $s \in \mathbb{C}^{K \times 1}$. The transmit signal is

$x = W s,$

and user $i$ receives

$y_i = h_i^H x + n_i = h_i^H W s + n_i,$

with vector form

$y = H W s + n.$

The per-user SINR and achievable rate are

$$\mathrm{SINR}_i = \frac{|h_i^H w_i|^2}{\sum_{j\neq i} |h_i^H w_j|^2 + \sigma^2},$$

and

$M$0

Representative optimization objectives include weighted sum-rate maximization under a total power constraint, per-antenna constraints, SINR balancing, and QoS-constrained power minimization (Xia et al., 2020).

The paper contrasts two architectural families. A data-driven BEAM-Net is a conventional CNN or DNN “black box” that maps inputs such as channel state information directly to outputs such as the beamforming matrix $M$1 without explicitly embedding domain knowledge. A model-driven BEAM-Net augments the NN with an explicit signal-processing module that can be placed before, within, or after the NN and whose parameters can be learned jointly. This module extracts key features, enforces constraints, and converts those features to the final precoder. The stated benefit is reduced dimensionality and improved accuracy, convergence, and generalization (Xia et al., 2020).

The classical baselines are iterative optimal or locally optimal solvers and low-complexity heuristics. The paper lists ZF and RZF as simple baselines and WMMSE, after Shi et al. (2011), as the standard iterative baseline for weighted sum-rate. WMMSE iteratively updates per-user linear receivers $M$2, MSE weights $M$3, and the precoder $M$4 to reduce the weighted sum-MSE, which is equivalent to maximizing weighted sum-rate under appropriate mappings. This baseline remains accurate but incurs iterative latency.

3. Learning strategies, complexity reduction, and deployment in wireless BEAM-Net

The wireless framework supports supervised, unsupervised or self-supervised, and hybrid learning. In supervised learning, labels are produced by iterative solvers such as duality-based SINR balancing algorithms, WMMSE, or QoS-constrained power minimization routines. A typical regression loss is

$M$5

or, when predicting key features such as the uplink power vector, a feature-space loss such as $M$6. In unsupervised learning, the original objective itself is used as a differentiable loss, for example

$M$7

for weighted sum-rate, or $M$8 for SINR balancing. The paper also describes a two-stage hybrid strategy: supervised pre-training using WMMSE-generated solutions followed by unsupervised fine-tuning with the task objective as loss (Xia et al., 2020).

A central model-driven idea is to predict low-dimensional key features instead of the full precoder. The example emphasized in the paper is the uplink power allocation vector for the dual uplink SINR-balancing problem, followed by a signal-processing conversion layer implementing uplink–downlink duality to recover the constrained downlink precoder. This reduces output dimension because the number of key features is much smaller than the $M$9 entries of $H \in \mathbb{C}^{K \times M}$0. Constraint handling is implemented through normalization or projection layers for total and per-antenna power constraints, including row-wise normalization for per-antenna limits (Xia et al., 2020).

The paper also discusses pruning, compression and weight sharing, and Huffman coding for common weights as NN-level reduction techniques. For generalization, it proposes training-set augmentation across multiple network sizes by padding inputs and outputs to a maximum size and sampling combinations of user and antenna counts with roughly equal probability. It also proposes transfer learning by replacing input and output layers as needed, inserting new intermediate layers, and applying different learning rates to transferred and newly added layers. Although simulations assume perfect CSI, robustness to pilot contamination, estimation errors, and domain shift is identified as an open direction (Xia et al., 2020).

Empirically, the simulations use downlink MISO with Rayleigh fading and large-scale path loss modeled as $H \in \mathbb{C}^{K \times M}$1 dB, with perfect CSI at the base station. For QoS-constrained power minimization under a QoS target of $H \in \mathbb{C}^{K \times M}$2 dB and $H \in \mathbb{C}^{K \times M}$3, BEAM-Net achieves lower required transmit power than ZF and outperforms the optimal iterative algorithm configured with a “high-threshold” convergence of $H \in \mathbb{C}^{K \times M}$4. The optimal algorithm with tighter threshold $H \in \mathbb{C}^{K \times M}$5 can further reduce power, but at the cost of much higher runtime, approximately $H \in \mathbb{C}^{K \times M}$6 slower than BEAM-Net. Feasibility exceeds $H \in \mathbb{C}^{K \times M}$7, and a generalized BEAM-Net trained via augmentation with $H \in \mathbb{C}^{K \times M}$8 and $H \in \mathbb{C}^{K \times M}$9 performs close to individually trained models (Xia et al., 2020).

The testbed demonstration addresses SINR balancing with per-antenna constraints using a 4-antenna transmitter formed by two NI USRP-2950 devices and four users emulated by two USRP-2950 devices. Under static channels, BEAM-Net outperforms ZF and RZF and is slightly inferior but close to the optimal iterative solution. Under dynamic channels with coherence time approximately $W \in \mathbb{C}^{M \times K}$0–$W \in \mathbb{C}^{M \times K}$1 ms, the optimal iterative solution degrades because computation finishes after CSI changes, while BEAM-Net yields significantly better BER than both ZF/RZF and the optimal iterative solver. The experiments therefore attribute practical advantage to low-latency inference rather than to absolute optimality under static CSI (Xia et al., 2020).

4. BEAM-Net in ultrasound beamforming

In ultrasound imaging, BEAM-Net addresses musculoskeletal bone imaging from single-plane-wave RF data. The problem setting emphasizes speckle noise, low spatial resolution, poor contrast, and anisotropic reflections, with bone characterized as a specular reflector. The paper states that when the transducer is tilted even modestly, receive channels remain misaligned after geometric delay compensation, causing destructive interference during summation. Conventional delay-and-sum beamforming with SPW acquisitions therefore trades image quality for frame rate, while multiple plane-wave compounding improves resolution and contrast but is computationally expensive and reduces frame rate (Madhusoodanan et al., 21 Jul 2025).

The input is one-channel SPW RF data indexed on a depth-lateral grid of $W \in \mathbb{C}^{M \times K}$2 samples per frame, acquired with a Verasonics Vantage 128 and an L11-5v linear array with 128 elements, center frequency $W \in \mathbb{C}^{M \times K}$3 MHz, and sampling rate $W \in \mathbb{C}^{M \times K}$4 MHz. The output is a beamformed, bone-enhanced B-mode image with improved cortical delineation, contrast, and speckle characteristics. The generator is a U-Net-like encoder–decoder with 3 downsampling encoder blocks and 3 upsampling decoder blocks with skip connections. The discriminator is a PatchGAN conditioned on a Bone Probability Map, with filters $W \in \mathbb{C}^{M \times K}$5, $W \in \mathbb{C}^{M \times K}$6 kernels, stride $W \in \mathbb{C}^{M \times K}$7, LeakyReLU activations, batch normalization, and a final sigmoid that produces a patchwise probability map (Madhusoodanan et al., 21 Jul 2025).

The Bone Probability Map is computed from CPWC images by combining local phase, feature symmetry, and integrated backscatter, normalized to $W \in \mathbb{C}^{M \times K}$8. It plays two roles. First, it is used to generate BEAM ground-truth images via an attention-weighted enhancement function,

$W \in \mathbb{C}^{M \times K}$9

with nominal attention weights $s \in \mathbb{C}^{K \times 1}$0, $s \in \mathbb{C}^{K \times 1}$1, and $s \in \mathbb{C}^{K \times 1}$2. Second, it conditions the discriminator, increasing adversarial sensitivity around bony structures (Madhusoodanan et al., 21 Jul 2025).

The receive DAS formulation for SPW is given by

$s \in \mathbb{C}^{K \times 1}$3

with

$s \in \mathbb{C}^{K \times 1}$4

For coherent plane-wave compounding over $s \in \mathbb{C}^{K \times 1}$5 steered transmissions, the paper defines transmit delay $s \in \mathbb{C}^{K \times 1}$6, receive delay $s \in \mathbb{C}^{K \times 1}$7, total delay $s \in \mathbb{C}^{K \times 1}$8, and compounded output

$s \in \mathbb{C}^{K \times 1}$9

BEAM-Net is explicitly intended to recover MPW-like structural benefits from SPW inputs (Madhusoodanan et al., 21 Jul 2025).

Training uses adversarial BCE loss and $x = W s,$0 reconstruction loss with $x = W s,$1, with the generator objective

$x = W s,$2

and discriminator BCE on conditioned inputs. Optimization uses Adam, learning rate $x = W s,$3, batch size $x = W s,$4, $x = W s,$5 epochs, and early stopping after $x = W s,$6 epochs without validation improvement. Training uses $x = W s,$7 synthetic plus $x = W s,$8 in-vivo wrist RF datasets with an $x = W s,$9 train/validation split, and testing uses $i$0 in-vivo wrist datasets from $i$1 volunteers and $i$2 synthetic wrist datasets from $i$3 subjects. Inference is approximately $i$4 ms per SPW frame at $i$5 cm depth, enabling approximately $i$6 frames per second (Madhusoodanan et al., 21 Jul 2025).

5. Metrics, experiments, and reported performance in ultrasound BEAM-Net

The ultrasound paper introduces the Edge Preservation Index as a new region-focused metric for evaluating structural fidelity in bone-enhanced ultrasound images. After Laplacian high-pass filtering to obtain $i$7 and $i$8, zero-mean high-pass signals are defined as

$i$9

and

$y_i = h_i^H x + n_i = h_i^H W s + n_i,$0

The evaluation suite also includes Contrast Ratio,

$y_i = h_i^H x + n_i = h_i^H W s + n_i,$1

Signal-to-Noise Ratio,

$y_i = h_i^H x + n_i = h_i^H W s + n_i,$2

Speckle Similarity Index based on histogram intersection, and SSIM with $y_i = h_i^H x + n_i = h_i^H W s + n_i,$3. Bone boundaries were manually segmented for in-vivo data, while synthetic data used expert annotations; background masks were generated via morphological dilation and subtraction (Madhusoodanan et al., 21 Jul 2025).

Against SPW-DASB on in-vivo wrist data, the reported table values are CR $y_i = h_i^H x + n_i = h_i^H W s + n_i,$4 and SNR $y_i = h_i^H x + n_i = h_i^H W s + n_i,$5, corresponding to relative increases of $y_i = h_i^H x + n_i = h_i^H W s + n_i,$6 and $y_i = h_i^H x + n_i = h_i^H W s + n_i,$7. Against SPW-DASB on synthetic wrist data, CR is $y_i = h_i^H x + n_i = h_i^H W s + n_i,$8 and SNR is $y_i = h_i^H x + n_i = h_i^H W s + n_i,$9, corresponding to increases of $y = H W s + n.$0 and $y = H W s + n.$1. The abstract summarizes overall improvements as “51.4–51% higher CR and 94.2–73.3% higher SNR” across in-vivo and synthetic datasets; the paper notes that the table-calculated improvements are higher for in-vivo and consistent for synthetic, and that differences likely arise from averaging over subjects or ROIs or from alternative baseline normalization in the abstract (Madhusoodanan et al., 21 Jul 2025).

Against MPW-DASB, the reported in-vivo values are CR $y = H W s + n.$2 versus $y = H W s + n.$3 and SNR $y = H W s + n.$4 versus $y = H W s + n.$5, corresponding to $y = H W s + n.$6 and $y = H W s + n.$7. On synthetic data, the reported values are CR $y = H W s + n.$8 versus $y = H W s + n.$9 and SNR $$\mathrm{SINR}_i = \frac{|h_i^H w_i|^2}{\sum_{j\neq i} |h_i^H w_j|^2 + \sigma^2},$$0 versus $$\mathrm{SINR}_i = \frac{|h_i^H w_i|^2}{\sum_{j\neq i} |h_i^H w_j|^2 + \sigma^2},$$1, corresponding to $$\mathrm{SINR}_i = \frac{|h_i^H w_i|^2}{\sum_{j\neq i} |h_i^H w_j|^2 + \sigma^2},$$2 and $$\mathrm{SINR}_i = \frac{|h_i^H w_i|^2}{\sum_{j\neq i} |h_i^H w_j|^2 + \sigma^2},$$3. The abstract summarizes these as “19.8–24.0%” improvements for in-vivo and “2.5–12.8%” for synthetic, with the paper remarking that the table-calculated improvements are in the same ballpark for synthetic and somewhat higher for in-vivo on the reported split (Madhusoodanan et al., 21 Jul 2025).

The comparison with other deep learning architectures includes U-Net, ResNet, and CNN baselines. On in-vivo data, BEAM-Net reports SSI $$\mathrm{SINR}_i = \frac{|h_i^H w_i|^2}{\sum_{j\neq i} |h_i^H w_j|^2 + \sigma^2},$$4, EPI $$\mathrm{SINR}_i = \frac{|h_i^H w_i|^2}{\sum_{j\neq i} |h_i^H w_j|^2 + \sigma^2},$$5, and SSIM $$\mathrm{SINR}_i = \frac{|h_i^H w_i|^2}{\sum_{j\neq i} |h_i^H w_j|^2 + \sigma^2},$$6, compared with ResNet values SSI $$\mathrm{SINR}_i = \frac{|h_i^H w_i|^2}{\sum_{j\neq i} |h_i^H w_j|^2 + \sigma^2},$$7, EPI $$\mathrm{SINR}_i = \frac{|h_i^H w_i|^2}{\sum_{j\neq i} |h_i^H w_j|^2 + \sigma^2},$$8, and SSIM $$\mathrm{SINR}_i = \frac{|h_i^H w_i|^2}{\sum_{j\neq i} |h_i^H w_j|^2 + \sigma^2},$$9. On synthetic data, BEAM-Net reports SSI $M$00, EPI $M$01, and SSIM $M$02, compared with ResNet values SSI $M$03, EPI $M$04, and SSIM $M$05. The averaged improvements versus ResNet are stated as CR $M$06, SNR $M$07, SSI $M$08, SSIM $M$09, and EPI $M$10. Paired Student’s $M$11-tests indicate $M$12 for BEAM-Net versus the next-best competitor across metrics (Madhusoodanan et al., 21 Jul 2025).

The ablation studies report that BEAM ground truth produces the best trained outcomes among GC, AHE, FBSR, and BEAM supervision, and that an attention-weight setting $M$13, $M$14, $M$15 yielded CR $M$16, SNR $M$17, SSI $M$18, and EPI $M$19 in-vivo. Direct application of analytical BEAM to SPW-DAS output underperforms BEAM-Net, with an approximately $M$20 drop in CR versus BEAM-Net. On noisy elbow datasets unseen during training, BEAM-Net reports CR $M$21, SNR $M$22, SSI $M$23, EPI $M$24, and SSIM $M$25, outperforming U-Net, ResNet, and CNN and revealing bone structures invisible in C-DASB (Madhusoodanan et al., 21 Jul 2025).

6. Other usages and nomenclature boundaries

A separate use of the label appears in heavy-ion phenomenology as Beam Energy dependence of Moments of net-charge. In that context, the object of study is the beam-energy dependence of event-by-event fluctuations of conserved quantities in Au+Au collisions at RHIC, not a neural network. The STAR Collaboration reports measurements of the net-charge moments—mean $M$26, variance $M$27, skewness $M$28, and kurtosis $M$29—for $M$30 and $M$31 GeV, with the standardized products

$M$32

mapped to QCD thermodynamic susceptibilities through

$M$33

Within present uncertainties, the products $M$34, $M$35, and $M$36 do not exhibit non-monotonic behavior versus $M$37, and the reported fluctuation measurements have been used to extract freeze-out temperatures $M$38–$M$39 MeV and baryon chemical potentials $M$40–$M$41 MeV across BES energies (Collaboration, 2014).

The optical paper "Beam Profiler Network (BPNet) -- A Deep Learning Approach to Mode Demultiplexing of Laguerre-Gaussian Optical Beams" is relevant chiefly as a boundary case. The authors consistently use BPNet, not BEAM-Net. BPNet is a two-network system comprising a calibration U-Net and a MobileNetV2 classifier for mode demultiplexing of Laguerre–Gaussian beams from intensity-only images. The classifier-only experiments on numerical data report $M$42 validation accuracy for one-, two-, and three-mode superpositions, while the full BPNet on held-out experimental test sets reports $M$43 single-mode accuracy when trained on single-mode numerics and $M$44 accuracy for two-mode superpositions when trained on matched two-mode numerics. Its inclusion clarifies that not every beam-related neural architecture in the literature uses the BEAM-Net label (Bekerman et al., 2019).

Taken together, these usages show that BEAM-Net is best understood as a context-sensitive label rather than a single established method. In wireless communications it denotes model-driven or data-driven neural beamforming; in ultrasound it denotes a specific adversarial SPW-to-B-mode reconstruction framework with bone enhancement attention; in heavy-ion work it is a shorthand for beam-energy-dependent net-charge fluctuation measurements; and BPNet remains a distinct optical method.