Degrees-of-Freedom Budgets for Sparse Ultrasound Arrays

Cheaper, lower-power ultrasound probes read out fewer receive channels. Removing elements spreads out the ones that remain, and wide spacing creates grating lobes: ghost copies of bright structures, placed where no tissue is. Deep networks can now take images from a 16-channel array and make them look like images from the full 128-channel array (Deep learning assisted sparse array ultrasound imaging). They cannot tell a clinician which parts of the output came from the measurements and which parts the network made up from what it saw in training. Wireless engineers faced a similar question when base stations grew to hundreds of antennas. Their answer was a count of effective degrees of freedom: the number of independent spatial patterns an array can actually tell apart. That count stops growing once the elements are dense enough, and the ceiling is set by the size of the aperture and the geometry of the scene, not by how many antennas are installed.

We propose to compute this count for the operator that maps tissue reflectivity in a chosen image region to the probe's channel data, and to use it for two jobs. The first is to predict, for each region and depth, how many receive elements are needed before image quality stops improving. The second is to split any network reconstruction into a part the sparse measurements support and a part supplied by the network's learned expectations, which yields a per-pixel map of possible hallucination. The contribution is not "information theory for ultrasound" in general. It is a specific, checkable budget that says where sparse-array deep reconstruction is trustworthy.

Core Claim

Near-field wireless analysis shows that effective degrees of freedom saturate once the array samples the field finely enough. The saturation point depends on aperture and on the angles the scene spans, not on a universal half-wavelength rule (Analytical Framework for Effective Degrees of Freedom in Near-Field XL-MIMO; Performance Analysis for Near-Field MIMO: Discrete and Continuous Aperture Antennas). Related work on bilateral near-field channels finds that the spatial bandwidth, and therefore the sampling density a region actually requires, depends on where the region sits relative to the array (Spatial Bandwidth of Bilateral Near-Field Channels for Linear Large-Scale Antenna Array System).

Our claim is that the same structure governs ultrasound receive arrays, with one extension. Tissue that bends or aberrates the acoustic wavefront changes the saturation point, and a homogeneous-medium formula cannot capture that change. Two things follow from the claim. First, the element count at which a sparse-array image stops improving should track the effective degrees of freedom of the region's own operator, not the probe-wide half-wavelength pitch rule. Second, the network error that cannot be removed should concentrate in image components outside the subspace the sparse array measures well.

Let $H_\Omega$ be the linear operator from reflectivities on a pixel grid in region $\Omega$ to receive-channel samples for a fixed transmit, and let $R = H_\Omega^{\mathsf H} H_\Omega$. Following the definition used by Wang et al., the effective degrees of freedom are

$$ \psi(\Omega) = \frac{\left(\operatorname{tr} R\right)^2}{\lVert R \rVert_F^2}. $$

This quantity stays near 1 when a single spatial pattern dominates. It approaches the rank of $R$ when all patterns are measured equally well.

For a reconstruction $\hat{x}$, let $P_s$ project onto the right singular vectors of the sparse-array operator whose singular values exceed the noise floor. We define the prior fraction of region $\Omega$ as

$$ \rho(\Omega) = \frac{\lVert (I - P_s)\,\hat{x}_\Omega \rVert^2}{\lVert \hat{x}_\Omega \rVert^2}. $$

This is the share of the output's energy that the sparse measurements cannot have supplied. The idea borrows the wireless observation that estimation accuracy saturates at the rank of the physical channel rather than at the array size (The Impact of Physical Channel on Performance of Subspace-Based Channel Estimation in Massive MIMO Systems; Massive MIMO: How many antennas do we need?). It turns that observation into a statement about which pixels a network is allowed to claim it recovered.

Research Program

Operators and degrees-of-freedom maps. We first build $H_\Omega$ for plane-wave transmits on a linear probe, over tiled regions spanning the image. We build it two ways. The first uses the homogeneous Green's function, the textbook model of a spherical wave spreading from a point. The second estimates the operator column by column from k-wave simulations that include layered speed-of-sound aberration and attenuation. For each region we compute $\psi$ as a function of how many elements are kept and how they are arranged.

The wireless analyses supply the method: spherical-wave Green's functions and the degrees-of-freedom ratio. They also mark the gap. Their media are lossless and homogeneous, so the aberrating version is new here and addresses a failure documented in ultrasound. CNN-Based Image Reconstruction Method for Ultrafast Ultrasound Imaging reports a performance drop on in vivo carotid data that the authors attribute to unmodeled sound-speed variation. LOTUS: Learning to Optimize Task-based US representations likewise lists the gap between simulated and real ultrasound as an open problem.

Constraint from mutual coupling. Wang et al. find that packing elements closer than half a wavelength eventually lowers degrees of freedom through electrical coupling between neighbors. We therefore never densify real hardware past the physical probe pitch; all our physical configurations are subsets of existing elements. The only denser-than-probe array we use is a virtual oversampled array of the kind Perdios et al. use as ground truth, and we use it only as the continuous-aperture reference, the upper limit on degrees of freedom.

Saturation prediction. For each region we record how image quality rises as receive elements are added. We then test whether the element count at which quality reaches 95% of the full-array value matches the count at which $\psi$ reaches 95% of its full-array value.

Prior-fraction maps. For trained sparse-to-dense restoration networks, we compute $\rho$ per region. We then ask whether it predicts per-pixel error and whether it flags structures that are present in the output but absent from the object.

Element selection. As a practical consequence, we choose the 16 or 32 retained elements to maximize the average $\psi$ over the regions of clinical interest. We compare this choice against uniform decimation, which is the layout Qi et al. use.

Baselines

Naive reference. The do-nothing point is delay-and-sum beamforming on uniformly decimated channels. Delay-and-sum is the standard method: it shifts each channel by its travel time and adds the channels up. Running it on decimated data shows the raw grating-lobe damage every method must repair. It is run on exactly the same channel subsets as every other method.

Strongest learned restoration method. This is a U-Net restoration network trained to map sparse-array images to full-array images, following the robust U-Net of Deep learning assisted sparse array ultrasound imaging. On ex vivo porcine teeth with 16 of 128 channels, that paper reports SSIM improving from 0.56 to 0.68 and lateral resolution of 0.15 mm, matching the full array.

The closest reconstruction competitor is the backprojection-plus-CNN pipeline of Perdios et al. That paper reports grating lobes reduced by more than 18 dB and runtimes of 22–39 ms per frame.

We found no reported result on Lacuna for learned restoration of channel-decimated PICMUS or CUBDL data. We will therefore retrain both architectures ourselves on identical training splits, with the same GPU budget and the same hyperparameter search budget.

Our method is not meant to beat these networks at image quality. It must beat them at knowing where they are wrong. Our element selection must beat uniform decimation at equal channel count. We will call element selection a win only if a network trained on the selected layout exceeds the uniform-layout network in SSIM and contrast-to-noise ratio, beyond seed-to-seed spread.

Rules for predicting saturation. Three simpler predictors compete with $\psi$. The first is the global half-wavelength pitch rule. The second is a geometric rule: grating lobes are harmless as long as they fall outside the angle the region subtends at the array. The third is the homogeneous-Green's-function $\psi$, which serves as an ablation isolating the contribution of the aberration-aware operator.

Uncertainty baselines. For hallucination detection we compare $\rho$ with pixel-wise variance across a deep ensemble and with Monte Carlo dropout, which estimates uncertainty by randomly switching off network units at test time. All are applied to the same trained networks.

We also run the artificial-input test of Overfit detection method for deep neural networks trained to beamform ultrasound images. That test feeds a network all-zeros, all-ones, or noise inputs to detect whether it has memorized its training data. It flags a network as a whole, not individual pixels, so it serves as a coarse reference.

Ablation of the key ingredient. This replaces the physics-defined projection $P_s$ with a projection onto random subspaces of matched dimension. It checks that $\rho$'s predictive power comes from the measurement operator and not from generic image statistics.

Why not just…?

Why not just use the textbook rules, half-wavelength pitch and lateral resolution proportional to wavelength times depth over aperture? In a homogeneous medium, the geometric grating-lobe rule may predict saturation nearly as well as $\psi$. This is a real threat to the saturation half of the proposal. The saturation experiment settles it by comparing prediction error for the two rules directly. If the geometric rule matches $\psi$ in homogeneous media and also in aberrating media, we will report that degrees-of-freedom analysis adds nothing for element-count design. The proposal would then rest on the prior-fraction maps, which no pitch rule can produce.

Why not just use ensembles or dropout to find hallucinations? Ensemble members trained on the same data share the same learned expectations. They can therefore agree confidently on a structure that no measurement supports, which is exactly the failure that memorizing beamformers show under Kukka's artificial inputs. The prior fraction $\rho$ is defined by the physics of the measurement and does not depend on the network. The hallucination experiment below compares the two by area under the ROC curve on structures planted to be absent. If ensembles match $\rho$, the cheaper option wins and we will say so.

Why not just compound many steered plane waves on a full probe? Coherent compounding fixes grating lobes, but Perdios et al. note that matching synthetic-aperture quality this way takes up to 192 acquisitions per frame. That cuts frame rate by about two orders of magnitude and raises power draw, which is unacceptable for ultrafast imaging and for battery-powered probes with few channels. We include full-probe compounding only as an upper reference for image quality.

Why not just rely on robust downstream segmenters? Segmentation networks such as Echo-DND score well on clean benchmarks. They work on the final image, though, so they cannot distinguish a grating-lobe ghost that a restoration network cleaned up well from one it invented. Echo-DND also takes about 2 minutes per frame. A per-pixel support map is complementary to such segmenters rather than a replacement for them.

Proposed Experiments

Data. We will use full-channel RF data from PICMUS and CUBDL, which cover simulations, phantoms, in vivo carotid, and cardiac acquisitions. We decimate channels in software, which yields exactly paired sparse and full data with no second acquisition. We will add Field II and k-wave simulations with controlled aberration layers of increasing strength, plus a virtual oversampled array as the continuous reference. If the porcine dataset of Qi et al. can be obtained, it gives a direct reproduction check against their reported numbers.

Decisive check 1: saturation prediction. Across regions, depths, and aberration strengths, we compare the element count at which image quality saturates with the counts predicted by $\psi$, by the geometric rule, and by the half-wavelength rule. Image quality is measured by SSIM against the full array, lateral full-width at half-maximum on point targets, and cyst contrast-to-noise ratio. The claim passes if the aberration-aware $\psi$ has clearly lower prediction error than both rules in aberrating media.

Decisive check 2: hallucination detection. In simulation, we plant absent-but-typical structures: we train on phantoms that often contain cysts and test on objects without them. We then measure whether $\rho$ flags invented cysts better than ensemble variance does, using area under the ROC curve. Separately, we measure the Spearman correlation between $\rho$ and per-region error on in vivo PICMUS and CUBDL frames.

Decisive check 3: element selection. We test whether layouts chosen to maximize $\psi$ raise SSIM and contrast-to-noise ratio for retrained restoration networks at 16 and 32 channels, compared with uniform decimation, over at least three training seeds.

Negative evidence. Failure on check 2 would mean that the measurement subspace does not explain where networks go wrong, and we would drop the certification claim. Failure on check 1 alone would narrow the paper to a diagnostic tool.