Which Plaque Cell Types Can Ultrasound See?

Whether a carotid plaque ruptures and causes a stroke depends heavily on which cells it contains. Lipid-laden immune cells and smooth muscle cells that have changed their behavior tend to weaken a plaque. Matrix-producing smooth muscle cells tend to stabilize it. Today these cell populations can only be counted after the plaque is surgically removed and stained or sequenced. Ultrasound sees plaques in living patients without radiation or surgery, but it measures something different: echoes from boundaries between materials of different density and stiffness. The best direct test of ultrasound against histology, from 1998, found that a texture feature correlated only at $r = -0.50$ with the volume of soft plaque tissue. The authors judged that too weak for individual patients (Wilhjelm et al.). Nobody has asked the sharper question: which parts of a plaque's cell-type mix leave any acoustic trace at all, and which are invisible in principle?

We propose to answer that question by borrowing the central idea of cellular composition confounding from genomics. A bulk tissue measurement is a mixture: the signal equals the cell proportions weighted by each cell type's signature, plus noise. Two cell types with similar signatures cannot be told apart, so a shift between them is confounded. We treat the local acoustic properties of plaque tissue the same way. We will estimate an "acoustic signature" for each plaque cell type from surgical specimens that are both scanned and spatially sequenced. We will use that signature matrix to predict which composition directions ultrasound can recover. Then we will test on raw ultrasound echo data from patients before surgery whether those directions, and only those, are recovered. The result is a map of what non-invasive cellular assessment of plaques can and cannot deliver. It also gives a principled training target for any model that tries to deliver it.

Core Claim

Plaque cell composition is only partly identifiable from ultrasound, and the identifiable part can be predicted before any patient model is trained. Let $h(x)$ be the vector of cell-type proportions at wall location $x$. Let $a(x)$ be the local acoustic descriptors: sound speed, density, attenuation, and statistics of sub-resolution scatterers, which shape the frequency spectrum of the raw echo signal. We model

$$a(x) = S\,h(x) + \varepsilon(x),$$

where $S$ is the acoustic signature matrix, with one column per cell type, and $\varepsilon$ is residual variation. Only components of $h$ that lie in the well-conditioned row space of $S$ can be recovered from echoes, however good the image model is. Our claim has three parts. First, this row space is low-dimensional, probably two to four directions. Second, it contains information about cell state beyond the three classical tissue classes of lipid, fibrous tissue and calcium. Third, a model trained to predict only these directions from raw radio-frequency (RF) echoes will beat image-only deep learning on held-out patients.

The second part is the claim most likely to fail, and it is the one that matters. If the acoustically visible directions reduce to "how much lipid, how much calcium", then ultrasound cell deconvolution is just tissue-class grading under another name. The study is designed to detect that outcome and report it.

Relation to Prior Work

Two lines of work motivate the gap, and each reaches only half of it. Spatial deconvolution from microscopy predicts dense cell-type maps from resected slides. Hist2Cell predicts up to 40 cell types from H&E images with a graph-transformer (Hist2Cell). A lightweight model on pathology foundation-model features matches it at a fraction of the compute (Integrating Pathology Foundation Models and Spatial Transcriptomics). POLARIS adds layer-specific expression modeling and super-resolved composition from histology (Jiawen Chen et al.). All three need tissue on a glass slide. We reuse them only as label generators on the excised specimen, never as the clinical instrument.

On the ultrasound side, two papers matter. Wilhjelm et al. compared processed B-mode texture with tissue-class volumes from histology. B-mode is the standard grayscale video image, produced by log-compressing and smoothing the raw echoes. They did not go down to cell types, and they named the loss of raw RF information as a main limit. Kraft et al. built a full-wave simulator that turns digitized histology into realistic intravascular echo data. It solves only the forward direction, and it assigns acoustic properties per stain class rather than per cell type. Our signature matrix $S$ replaces their stain-class lookup with a fitted, cell-type-resolved mixture. Their simulator becomes our tool for predicting identifiability.

Generative neural physics shows that sound speed and density can be inverted quickly with a neural operator. Its acoustic-fluid assumption excludes attenuation, which we expect to carry much of the cell-type signal. That limitation constrains us: we cannot use it as the inversion engine for plaques.

From the genomics side, CSeQTL shows that naive linear models with interactions on log-transformed mixtures leak an abundant cell type's effect into rare ones. That result motivates our handling of the mixture model. We fit $S$ on untransformed acoustic quantities, and we test explicitly for leakage of tissue-class effects into cell-state predictions.

Research Program

The program has three stages.

The first stage builds a paired cohort. Patients scheduled for carotid endarterectomy receive a preoperative scan with raw RF capture, in the cross-sectional and longitudinal planes Wilhjelm used, calibrated to the blood pool. The excised specimen then follows two paths. First, it is scanned ex vivo at high frequency in a water bath, in the style of Kraft's in-vitro platform, at the same positions that are later sectioned. Second, it is sectioned for H&E. A subset of sections also undergoes spatial transcriptomics, deconvolved with cell2location against a plaque single-cell reference. The foundation-model regressor is then trained on that subset and extends composition labels to all H&E sections.

We plan for roughly 80 patients. That is enough to detect a held-out correlation of about 0.3 with reasonable power. Ex vivo scans of resected specimens support many thousands of matched patches, which is where $S$ is estimated.

The second stage estimates $S$ and its identifiable subspace. On the ex vivo data, the voxelwise match between echo and composition is nearly exact, because the scan and the sections come from the same registered specimen. We extract local acoustic descriptors and fit $S$ by constrained regression on the composition maps. The singular value spectrum of $S$, together with a noise level measured from repeated scans, defines the recoverable directions. We then rotate these directions into interpretable axes and ask whether any axis stays predictive after the lipid, fibrous and calcium fractions are projected out.

Kraft's simulator supplies a second, physics-only estimate. We forward-simulate sections whose cell composition we perturb along each candidate direction. If a perturbation does not change the simulated echo, that direction lies in the null space.

The third stage builds the clinical inverse model. A network takes patches of preoperative RF data and predicts only the coordinates of composition in the identifiable subspace. It is pretrained on simulated echoes from the full set of H&E sections and fine-tuned on in-vivo patches registered to histology. Registration uses the carotid bifurcation, calcium deposits, and the intima-media double-line as landmarks. The model's output layer mirrors the arterial wall's layers, so that intimal and medial composition are predicted separately. This addresses the shared-layer simplification that POLARIS acknowledges.

Baselines

The naive reference predicts every patient's composition as the cohort mean in each subspace coordinate. It defines zero skill. Any method must beat it on held-out patients in per-coordinate Pearson correlation and in L1 error on the reconstructed proportions.

The classical baseline is Wilhjelm's pipeline: first-order histogram and gray-level co-occurrence texture features on calibrated B-mode frames, fed into linear regression. We extend it to predict our subspace coordinates. It runs on exactly the same patients, folds and registered regions as our model. Its reported result in the file is GLCM Contrast versus relative soft tissue volume, $r = -0.50$ (Wilhjelm et al.). Matching that on the soft-tissue coordinate checks that our pipeline is sound, not that it is better.

The strongest image-only competitor is a modern convolutional segmentation backbone of the kind used by FRDD-Net (Yanhan Li et al.), whose test-set Dice of 83.65% shows that such networks extract plaque structure well from B-mode. We repurpose it as a regressor on B-mode patches with the same targets. It gets the same hyperparameter-search budget, the same augmentation, and the same simulated pretraining data, rendered as B-mode. No reported result was found on Lacuna for ultrasound-to-cell-type composition regression with this or any network, so this baseline's number will be produced here.

Beating this baseline means a statistically significant gain in held-out correlation on at least one non-tissue-class coordinate. We test this with a paired bootstrap over patients.

Two ablations isolate the key ingredients. The first trains our network on unrestricted 20-plus cell-type targets instead of the identifiable subspace. If restriction does not help on the identifiable coordinates, or the unrestricted model shows apparent skill on coordinates the analysis calls null, the identifiability analysis is wrong. The second removes simulated pretraining, which tests whether the physics simulator earns its cost.

As a ceiling rather than a competitor, we report the foundation-model histology regressor applied to the specimen's own H&E. Its numbers on breast cancer cohorts exceeded Hist2Cell, whose informative cell types reached $R > 0.80$ (Hist2Cell). It shows how much of the gap between invasive and non-invasive assessment remains.

Why not just…?

Why not just regress cell proportions from B-mode with a large network? This is the real threat. If the convolutional baseline matches our RF model on every identifiable coordinate, then raw echoes and physics add nothing, and we will say so. Wilhjelm's attribution of error to the information discarded by B-mode processing is a hypothesis, not a measurement. Our RF-versus-B-mode comparison, run with identical networks and pretraining, is the experiment that measures it.

Why not just predict lipid, fibrous and calcium fractions and map them to cell types with a fixed table? If plaque cell state is mostly determined by tissue class, this simpler route is as good as ours. The partial-correlation test answers it directly: after regressing out the three predicted tissue fractions, does any ultrasound-predicted coordinate still correlate with held-out sequencing-derived composition? We will also fit the lookup-table model explicitly as a baseline. If it ties, the paper's conclusion becomes that the acoustically visible cell composition is the tissue-class composition. That is a useful negative result and closes off a direction others might fund.

Why not just sequence the endarterectomy specimen? For patients who already have surgery, sequencing is better, and we use it as ground truth. The purpose here is the much larger population who are scanned but never operated on, where composition would inform whether to operate.

Why not use 3T MRI instead? MRI may well resolve more. However, the identifiability question is modality-specific, and ultrasound is the bedside tool already used for intima-media thickness and Doppler screening. The neural tomography work reached 1.09 mm against MRI's 0.88 mm in limbs (Generative neural physics), which suggests the resolution gap is not the deciding factor.

Proposed Experiments

The first decisive check is internal. Coordinates the signature analysis labels as null should show near-zero held-out correlation for every method, including the unrestricted ablation. Coordinates it labels as identifiable should show positive correlation in the ex vivo data. Disagreement falsifies the mixture model.

The second check is the partial-correlation test against tissue-class fractions on preoperative scans. It decides whether the work is about cells or only about tissue types.

The third compares RF against B-mode at matched capacity. It decides whether raw echoes matter.

We will also report two diagnostics. The first is leakage, in the CSeQTL sense: whether a strong calcium signal produces spurious predictions for rare immune-cell coordinates. We test this by shuffling the cell-state labels within strata of calcium fraction. The second is agreement between the identifiable subspace estimated from Kraft-style simulation and the one estimated from real ex vivo echoes. This shows whether simulation alone could replace costly paired cohorts in other vascular beds.

All results are leave-patient-out, with patient-level bootstrap intervals. Registration error is estimated on landmark positions held out from fitting and propagated into the reported intervals.