Bone Probability Map (BPM) Overview

• Bone Probability Map (BPM) is a spatial representation that assigns normalized probability scores to image regions, indicating the likelihood of bone presence.
• It is derived from beamformed ultrasound data using integrated backscatter, local phase, and feature symmetry to highlight cortical interfaces.
• BPM functions as a training prior and discriminator conditioning in adversarial models, improving structural fidelity and image enhancement outcomes.

A Bone Probability Map (BPM) is a spatial representation that encodes the likelihood that image locations correspond to bone or bone-related structure. In the most explicit arXiv formulation, BEAM-Net defines BPM as a normalized scalar map over image coordinates $(x,y)$ with values in $[0,1]$, where larger values indicate higher probability that the corresponding image location belongs to a bony region; it is computed in the beamformed ultrasound image domain and used simultaneously as a soft bone-likelihood image, an image-domain attention prior, and discriminator conditioning during adversarial learning (Madhusoodanan et al., 21 Jul 2025). Across adjacent literatures, the same functional role is often realized by related but non-identical constructs, including BBM occupancy masks, bone-only projection maps, probabilistic cortical shell models, signed distance fields, and intrinsic surface parameterizations. This suggests that BPM is not a universally standardized object, but a modality-dependent family of bone-localizing priors.

1. Conceptual definition and scope

Within BEAM-Net, BPM is explicitly constructed rather than learned as an internal latent feature map. It is derived from a bone-enhancement pipeline operating on a coherently compounded ultrasound image, and the paper states that “$BPM(x,y)$ results in a probability map that highlights bony regions” (Madhusoodanan et al., 21 Jul 2025). Its physical interpretation is the spatial likelihood of bone presence, especially around the bone-soft tissue interface and associated shadowing signatures. Its algorithmic interpretation is broader: the same map functions as a soft probability image, an attention prior during enhancement, and conditioning information for the discriminator.

Several negative definitions are equally important. In that formulation, BPM is not primarily a binary mask, although a thresholded version is mentioned for enhancement; it is not a generator output; it is not a segmentation annotation; and it is not an explicitly supervised prediction branch. This distinction matters because “probability map” in musculoskeletal imaging can refer either to a calibrated posterior over bone occupancy or, more loosely, to a soft spatial prior emphasizing likely osseous structure. The BEAM-Net usage belongs to the latter category (Madhusoodanan et al., 21 Jul 2025).

Work	Representation	Function
BEAM-Net (Madhusoodanan et al., 21 Jul 2025)	Soft normalized map in image space	Bone-likelihood prior, enhancement term, discriminator conditioning
Whole-body CE CT (Leydon et al., 2020)	BBM sigmoid output / binary mask	Subject-specific BBM occupancy prior
Hip X-ray to QCT (Gu et al., 2023)	PF-DRR / 2D BMD distribution	Bone-only projection map, not probability
Clinical QCT AbS (Reinhold et al., 2020)	Probabilistic cortical shell model	Boundary/material uncertainty along normals
High-order CT SDF (Besler et al., 2021)	Signed distance field	Geometric soft-occupancy surrogate
Bone surface mapping (Fan et al., 2010)	Inertial surface map	Registration and correspondence substrate

This broader landscape is useful because it prevents a common misconception: BPM is not a single canonical representation transferable unchanged across ultrasound, CT, radiography, and morphometric modeling. Rather, the term identifies a modeling objective—localized bone likelihood or bone-focused spatial prior—whose mathematical realization depends on image physics and downstream task.

2. Hand-crafted BPM construction in ultrasound beamforming

In BEAM-Net, BPM is generated from multi-plane-wave data rather than from the single-plane-wave data used at inference. First, MPW-RF data from 73 steered plane waves spanning $-18^\circ$ to $+18^\circ$ are beamformed and coherently compounded to obtain a CPWC image $I(x,y)$, which is then normalized to $I_{\text{norm}}(x,y)$. The pipeline then combines three classes of ultrasound-derived evidence: integrated backscatter, local phase, and feature symmetry (Madhusoodanan et al., 21 Jul 2025).

Integrated backscatter is computed columnwise from the normalized image and is motivated by the fact that bone is a strong reflector. A shadow map is also constructed using Gaussian-weighted averaging along image columns, with Gaussian window

$G(i) = e^{\frac{-i^2}{2\sigma^2}},$

where $\sigma$ controls the window size. The paper states that this smooths the image while preserving sharp transitions near bone boundaries. It also notes an implementation-level omission: the shadow map is described conceptually, but no final explicit equation is given for how it enters the BPM formula.

Local phase-based features are extracted using an Analytic Estimator with a multi-scale set of log-Gabor filters. The features are local phase, feature symmetry, and local energy, although only local phase and feature symmetry explicitly appear in the final BPM equation. The final normalized map is

$$BPM(x, y) = \frac{\left( LP(x, y) \times FS(x, y) \times \left[ 1 - IBS(x, y) \right] - \text{min} \right)}{\text{max} - \text{min}},$$

where $[0,1]$0 and $[0,1]$1 are the minimum and maximum of the product term over the image. This normalization yields a soft-valued image in $[0,1]$2 (Madhusoodanan et al., 21 Jul 2025).

The BPM is then fused with the CPWC image to generate the bone-enhanced target:

$[0,1]$3

with $[0,1]$4, $[0,1]$5, and $[0,1]$6 in the proposed setting. A thresholded BPM using Otsu’s method is mentioned in the text, but the enhancement equation writes $[0,1]$7 directly. The paper therefore leaves a small ambiguity: BPM is definitively a soft probability map, but a thresholded or binarized version apparently participates in enhancement, and the exact relation between the two is not specified beyond that sentence (Madhusoodanan et al., 21 Jul 2025).

This formulation embodies the paper’s bone physics argument. Local phase emphasizes structural transitions, feature symmetry captures asymmetry associated with specular bone reflections, and integrated backscatter captures strong reflected energy and shadow-related behavior. The resulting BPM is therefore a composite likelihood field tuned to cortical interfaces rather than a generic brightness map.

3. BPM as training prior rather than network output

BPM is not the direct supervision target for BEAM-Net. Instead, it is used to create the BEAM image, and the BEAM image becomes the ground-truth target during training. The ground-truth generation pipeline is: acquire MPW-RF data; beamform and compound into a CPWC image; compute BPM from CPWC via IBS, LP, FS, and normalization; fuse the CPWC image and BPM through the weighted enhancement equation to obtain the BEAM image; and train BEAM-Net to map SPW-RF input to this BEAM image (Madhusoodanan et al., 21 Jul 2025).

Architecturally, BPM enters through the discriminator rather than the generator. The generator is a U-Net with 3 encoder layers, 3 decoder layers, and skip connections; its input is SPW-RF data of size $[0,1]$8, and it outputs a reconstructed B-mode image. No BPM is concatenated to the generator input, and no BPM prediction branch is described. The discriminator is conditioned on BPM through a two-channel input consisting of either a real ground-truth image or a generated image together with the BPM of that image. The paper states that this conditioning channel provides contextual information that helps the discriminator distinguish real from fake images more effectively (Madhusoodanan et al., 21 Jul 2025).

This is why the “Bone Enhancement Attention Mechanism” should not be interpreted as a standard learned self-attention or feature-attention block. There are no equations showing BPM-based modulation of skip connections, hidden feature tensors, or generator outputs, and there is no separate BPM loss. The adversarial objective is standard binary cross-entropy, and the generator minimizes an adversarial term plus an $[0,1]$9 reconstruction term with $BPM(x,y)$0. The phrase that BPM “acts as an attention mechanism to enforce higher structural similarity around bony regions” is therefore operational rather than architectural: BPM shapes the training target and the discriminator’s decision context, but it is not learned as an explicit feature map inside the generator (Madhusoodanan et al., 21 Jul 2025).

A related practical point is that BPM is required only during training. Training uses triplets $BPM(x,y)$1, but deployment uses the trained generator alone. The paper reports inference in under about $BPM(x,y)$2 ms, without MPW compounding or online BPM computation, and contrasts this with about $BPM(x,y)$3 s for MPW beamforming and compounding in the reported setup (Madhusoodanan et al., 21 Jul 2025).

4. Quantitative behavior and structural evaluation

The experimental evidence for BPM is indirect, because no ablation removes BPM alone while keeping the rest of the system fixed. There is no reported “BEAM-Net without BPM” variant and no comparison against an alternative conditioning prior within the same architecture. Exact isolated attribution is therefore unavailable. What the paper does provide are architecture ablations, target-generation ablations, and enhancement-weight ablations that all depend on the BEAM/BPM formulation (Madhusoodanan et al., 21 Jul 2025).

BEAM-Net is compared against conventional DASB and existing deep learning architectures using Contrast Ratio, Signal-to-Noise ratio, Speckle Similarity Index, Structural Similarity Index, and the newly introduced Edge Preservation Index. Relative to SPW-DASB, it shows $BPM(x,y)$4-$BPM(x,y)$5 higher CR and $BPM(x,y)$6-$BPM(x,y)$7 higher SNR on in-vivo MSK and synthetic RF datasets. Relative to MPW-DASB, it yields $BPM(x,y)$8-$BPM(x,y)$9 improvements in CR and SNR on in-vivo MSK data and $-18^\circ$0-$-18^\circ$1 improvements on synthetic data. On in-vivo data, the reported BEAM-Net metrics are $-18^\circ$2, $-18^\circ$3, $-18^\circ$4, $-18^\circ$5, and $-18^\circ$6; on synthetic data they are $-18^\circ$7, $-18^\circ$8, $-18^\circ$9, $+18^\circ$0, and $+18^\circ$1 (Madhusoodanan et al., 21 Jul 2025).

The EPI is especially relevant because BPM is intended to improve structural fidelity at cortical boundaries. It is defined as

$+18^\circ$2

where

$+18^\circ$3

Here $+18^\circ$4 and $+18^\circ$5 are high-pass filtered versions of the ground-truth and predicted images using a standard $+18^\circ$6 Laplacian approximation. BPM does not appear explicitly in the metric, but it influences EPI indirectly through target construction and discriminator conditioning (Madhusoodanan et al., 21 Jul 2025).

The target-generation ablation is the closest BPM-related comparison. Using the same PatchGAN-based architecture, the BEAM targets outperform Gamma Correction, Adaptive Histogram Equalization, and Frequency-Based Super-Resolution. The enhancement-weight ablation further shows that BPM contribution matters quantitatively; the best reported setting is around $+18^\circ$7, $+18^\circ$8 or $+18^\circ$9, and $I(x,y)$0. Conversely, direct analytical BEAM enhancement applied to SPW beamformed data remains visibly inferior, with CR dropping by about $I(x,y)$1 relative to BEAM-Net. This indicates that BPM-based analytical enhancement alone is insufficient; the learned network is needed to reconstruct subtle bone detail from SPW-RF data (Madhusoodanan et al., 21 Jul 2025).

5. BPM-like formulations in CT, radiography, and physics-based modeling

Outside BEAM-Net, the term BPM is often absent even when the underlying construct is closely analogous. In low-dose, contrast-enhanced whole-body CT, the closest equivalent is a subject-specific BBM occupancy prior derived from a 2D U-net that segments combined bone-bone marrow regions rather than cortical bone alone. That distinction is crucial: if such outputs are averaged into a population map, the result is a BBM occupancy prior, not a pure cortical bone prior. The method achieves mean Dice coefficients of $I(x,y)$2, $I(x,y)$3, and $I(x,y)$4 on two internal and one external dataset, uses wider HU clipping ranges such as $I(x,y)$5 to $I(x,y)$6 HU or $I(x,y)$7 to $I(x,y)$8 HU, and selects post-sigmoid thresholds from the precision-recall curve rather than fixing $I(x,y)$9 (Leydon et al., 2020). This suggests a BPM pipeline in CT can be built from stable segmentation masks, but those masks remain binary and uncalibrated in a probabilistic sense.

In plain radiography for opportunistic osteoporosis screening, the closest BPM analogue is the proximal femur region DRR. The network learns decomposition into projections of bone-segmented CT/QCT and produces a PF-DRR that the paper characterizes as a pixel-wise BMD estimation and a 2D BMD distribution of the proximal femur bone. This is not a bone-probability map: it is a continuous projected density or attenuation-like image. Its value is that it localizes bone contribution while suppressing irrelevant soft-tissue content, and it supports BMD estimation with reported Pearson correlation coefficients of $I_{\text{norm}}(x,y)$0 for DXA-measured BMD and $I_{\text{norm}}(x,y)$1 for QCT-measured BMD in the best decomposition-guided regression setting (Gu et al., 2023).

Clinical QCT cortical modeling goes further from occupancy maps and closer to latent anatomy. The analysis-by-synthesis method of cortical thickness estimation defines a probabilistic generative cortical shell model rather than a voxelwise BPM. Uncertainty is parameterized through cortical half-width $I_{\text{norm}}(x,y)$2, center offset $I_{\text{norm}}(x,y)$3, and compartment densities $I_{\text{norm}}(x,y)$4, all fitted from blurred profile observations using MAP estimation and MCEM. The result is a surface-based thickness map with strong agreement to HR-pQCT, including $I_{\text{norm}}(x,y)$5 and root mean square error below $I_{\text{norm}}(x,y)$6, whereas standard QCT apparent thickness overestimates cortical thickness by $I_{\text{norm}}(x,y)$7 and shows $I_{\text{norm}}(x,y)$8 (Reinhold et al., 2020). The nearest BPM interpretation here is a probabilistic boundary-and-material model along surface normals.

A different geometric route appears in high-order signed distance fields from CT. That work constructs a high-order signed distance field $I_{\text{norm}}(x,y)$9 directly from density-calibrated two-phase CT data, yielding sub-voxel interface localization free of the quantization artifact associated with binary distance transforms. The paper does not present a statistically calibrated BPM, but it explicitly provides regularized Heaviside and signed distance machinery from which a soft occupancy field can be derived. A plausible implication is that $G(i) = e^{\frac{-i^2}{2\sigma^2}},$0 can function as a smooth bone-occupancy surrogate when sub-voxel geometry matters more than binary labeling (Besler et al., 2021).

In ultrasound delineation constrained by propagation physics, the explicit output is again not a BPM but a contour. However, the factor-graph formulation produces BPM-like ingredients: unary probabilities for soft tissue and acoustic shadow, confidence and shadowing maps, and a phase-symmetry-modulated $G(i) = e^{\frac{-i^2}{2\sigma^2}},$1 transition likelihood that identifies the most plausible bone boundary. This is a physics-informed surrogate for local bone probability, with average RMSE $G(i) = e^{\frac{-i^2}{2\sigma^2}},$2 mm, symmetric Hausdorff distance $G(i) = e^{\frac{-i^2}{2\sigma^2}},$3 mm, and detection of $G(i) = e^{\frac{-i^2}{2\sigma^2}},$4 of annotated bone surfaces (Ozdemir et al., 2020).

Finally, surface probability modeling presupposes correspondence. The inertial surface mapping technique for bone in vivo does not estimate bone likelihood, but it establishes a repeatable, anatomy-intrinsic coordinate system via principal axes of inertia, defines a prime meridian, and develops the closed bone surface into a map-like representation (Fan et al., 2010). A plausible implication is that any surface-based BPM over repeated or population scans requires this kind of normalization before probabilities can be meaningfully compared across time or subjects.

6. Ambiguities, limitations, and interpretive boundaries

The principal ambiguity in BEAM-Net concerns implementation detail rather than concept. BPM is clearly defined as a normalized soft map in the image domain, yet the enhancement stage mentions a thresholded BPM using Otsu’s method while the published enhancement equation retains the soft-valued $G(i) = e^{\frac{-i^2}{2\sigma^2}},$5. The paper also describes a shadow map but does not provide a final explicit equation showing how that shadow map enters the BPM formula. These omissions do not change the conceptual role of BPM, but they do constrain exact reproducibility (Madhusoodanan et al., 21 Jul 2025).

A second limitation is attribution. Because BPM is embedded in both target generation and discriminator conditioning, the paper does not isolate how much of the reported gain is due to the prior itself as opposed to the adversarial architecture or the BEAM target formulation more generally. Sensitivity to BPM estimation error is also not systematically analyzed; the closest available evidence is the enhancement-weight ablation, which shows that performance changes with the weighting of the BPM term (Madhusoodanan et al., 21 Jul 2025).

Training-time data requirements also matter. BPM construction in BEAM-Net requires MPW-RF acquisitions during training in order to generate CPWC images and BEAM targets, even though inference later uses SPW-RF only. The paper further notes limited access to RF data from commercial POCUS systems and a relatively small dataset, and proposes larger clinical studies and extension to other MSK anatomies such as hip and shoulder (Madhusoodanan et al., 21 Jul 2025).

Across modalities, related representations impose their own interpretive boundaries. The CE CT method produces BBM masks rather than calibrated probabilities and is limited by slice-wise modeling and restricted internal scanner diversity (Leydon et al., 2020). The PF-DRR is a bone-focused intermediate map, but it is a continuous projection image rather than $G(i) = e^{\frac{-i^2}{2\sigma^2}},$6 (Gu et al., 2023). The cortical AbS model is probabilistic, but its uncertainty is attached to latent profile parameters rather than to a dense voxel map (Reinhold et al., 2020). The high-order SDF is geometric rather than probabilistic, and any conversion to a BPM remains a downstream construction rather than a validated output of the method (Besler et al., 2021).

Taken together, these distinctions support a narrow technical conclusion. A BPM is most rigorously defined when it is an explicitly normalized soft bone-likelihood map in image space, as in BEAM-Net. More broadly, the term also serves as a useful umbrella for modality-specific bone priors that localize osseous structure, encode uncertainty or soft occupancy, and guide enhancement, segmentation, or morphometry. The literature therefore treats BPM less as a single standardized data type than as a class of bone-aware spatial representations adapted to the physics, sampling geometry, and supervision regime of each imaging problem.