Ultrasound scanners build an image by assuming that sound travels through the body at one fixed speed, usually 1540 m/s. Real tissue does not cooperate. Fat is slower, and glands and tumors are faster. Every mismatch between the assumed speed and the true speed blurs and shifts the image. Radiomics, the practice of turning an image region into hundreds of numerical texture and shape descriptors, then reads that blur as if it were tissue structure.
This matters most when a clinician compares two scans of the same patient taken weeks apart, for example to judge whether a breast tumor is shrinking under chemotherapy. Between visits, the tissue the sound passes through can change: the overlying fat thins, the tumor softens, the probe sits at a different angle. When that path changes, the blur pattern changes too, and some texture descriptors will drift even if the lesion is unchanged. We propose to measure how much of a longitudinal radiomic change is a sound-speed artifact. We also propose a fix: estimate the sound-speed maps of both visits jointly, with an explicit prior that the medium stays the same except where the tissue truly changed, and only then compute radiomic differences.
Core Claim
The claim has two parts. The first is diagnostic. In ultrasound, the visit-to-visit difference in standard radiomic features contains an artifact term driven by changes in sound-speed mismatch along the acoustic path. For many second-order texture features, this term can be comparable to the biological change the clinician wants to detect. Second-order features here means statistics of how neighboring pixel intensities relate, such as gray-level co-occurrence and run-length matrices.
The second part is methodological. Correcting each visit independently with an existing sound-speed estimator is not enough. Each estimate has its own error, roughly 10 m/s for the best learned estimator on simulated data, and those errors differ between visits, so independent correction can trade one artifact for another. A joint, change-constrained estimate should suppress the artifact term more than either uncorrected imaging or independent per-visit correction.
Longitudinal reasoning is the application setting, reproducibility is the evaluation lens, and radiomics is the quantity being protected.
Where This Sits
The sound-speed literature supplies the estimators. Iterative Pulse–Echo Tomography for Ultrasonic Image Correction inverts phase shifts in echo signals to recover slowness maps. Slowness is the reciprocal of sound speed. It feeds each estimate back into the beamformer, the step that converts raw echoes into an image, and reports up to 18.2% higher coherence factor than single-pass inversion. Investigating Pulse-Echo Sound Speed Estimation in Breast Ultrasound with Deep Learning trains an encoder-decoder on k-Wave simulations, a standard acoustic wave simulator, to map three plane-wave acquisitions to a sound-speed map. It reports a mean absolute error of 10.3 ± 5.60 m/s in simulation and degrades beyond 30 mm depth. Probing the speckle to estimate the effective speed of sound shows that the statistics of speckle, the grainy interference pattern in ultrasound images, carry sound-speed information.
All three papers evaluate a single time point with image-quality metrics. None asks what their errors do to texture descriptors, or to differences between two scans. We reuse them as components and as baselines.
The reproducibility and radiomics work motivates the evaluation. Fréchet Radiomic Distance (FRD) compares datasets using 464 radiomic features and reports an out-of-distribution detection AUC of 0.94 versus 0.79 for FID. It remains stable with only 10–100 images. However, it was validated only on CT and MRI, and its authors leave ultrasound and temporal data as future work. Both are exactly the regime here. The breast ultrasound synthesis study Hybrid Diffusion Model for Breast Ultrasound Image Augmentation still scores speckle realism with FID, which illustrates that the field lacks a speckle-aware yardstick.
The longitudinal work constrains the design. DeltaNet subtracts current from prior chest X-ray features. CoCa-CXR shows that unrestricted comparison spreads attention onto unrelated anatomy, and that restricting comparison to matched regions helps. It reports 65.0% progression accuracy on MS-CXR-T. Both papers assume the imaging physics is the same at both visits. In X-ray that is roughly true. In ultrasound it is not, and that is the gap we target. We adopt CoCa-CXR's lesson that comparison should be regional, but we apply it at the level of the acoustic medium rather than learned features.
Research Program
The first component is a controlled longitudinal benchmark, because the file contains no public longitudinal ultrasound dataset with raw channel data. We generate paired "visits" in k-Wave, following the Gaussian-random-field medium construction used by Simson et al. We control two factors independently.
The first factor is biological change inside a lesion region. It covers volume shrinkage, internal scatterer density change, and change in the lesion's own sound speed.
The second factor is nuisance change along the acoustic path. It covers thickness and speed of an overlying fat layer, small probe translations and tilts, and a global speed offset.
Each pair has a known ground truth: progressed, regressed, or stable lesion, crossed with a stable or changed path. We complement the simulations with physical phantoms in the style of the agar-ham phantom of Zengqiu et al. In these, a fixed inclusion is imaged twice, once under its original overlying layer and once after that layer is swapped for one of different speed. This gives a real-scanner "stable lesion, changed path" condition.
We also use the simulated cylinders portion of OADAT as a second, optoacoustic modality with known ground truth. Optoacoustic imaging uses laser-induced sound and shares the constant-speed reconstruction assumption.
The second component is the measurement. For each radiomic feature $f$, we define a nuisance-to-signal ratio:
$$ R_f = \frac{\operatorname{median}\, |\Delta f|\ \text{over pairs with path change only}}{\operatorname{median}\, |\Delta f|\ \text{over pairs with biological change only}} $$
Here $\Delta f$ is the visit-2 minus visit-1 value of feature $f$ in the registered lesion region. $R_f$ near or above 1 means feature $f$ cannot tell a changed path from a changed tumor. We report $R_f$ across the full family of first-order, shape, and texture features, including filtered variants.
We also extend FRD to visit pairs. We compute it between the distribution of radiomic deltas from stable pairs and from changed pairs, using features computed on envelope-detected ultrasound with explicitly reported settings.
The third component is joint estimation. Let $c_1$ and $c_2$ be the sound-speed maps of the two visits, $y_1$ and $y_2$ the raw echo data, $\mathcal{L}$ the data-fit term of an existing estimator, and $T$ a deformation that aligns visit-2 anatomy to visit-1 anatomy. We estimate both maps together:
$$ \min_{c_1, c_2, T}\ \mathcal{L}(y_1 \mid c_1) + \mathcal{L}(y_2 \mid c_2) + \lambda \,\| c_1 - c_2 \circ T \|_1 $$
The data-fit term $\mathcal{L}$ can be iterative CUTE's phase-shift misfit or a learned estimator's output used as a soft prior. The $L_1$ penalty, weighted by $\lambda$, encodes the belief that the medium changed only in a sparse set of locations. It plays the role, at the level of the acoustic medium, that regional masking plays in CoCa-CXR.
Both visits are then re-beamformed with their jointly estimated maps before radiomics are extracted. The estimated speed difference $c_1 - c_2 \circ T$ inside the lesion is kept as an extra longitudinal feature, because tumor sound speed is itself a candidate biomarker.
The fourth component is a deliberately simple longitudinal task. A logistic-regression or small gradient-boosted classifier labels each pair as progressed, regressed, or stable from radiomic deltas. The classifier is kept fixed across all imaging pipelines, so differences reflect the images and not the model.
Baselines
The naive reference is delay-and-sum beamforming at 1540 m/s for both visits, followed by the identical feature extractor and classifier. Delay-and-sum is the standard beamformer, which aligns echoes using the assumed constant speed. This is what a clinical radiomics study would do today. It defines the artifact level we are trying to remove.
The cheapest physics-aware baseline is per-visit global speed autofocus. Each visit is beamformed at the single uniform speed between 1420 and 1600 m/s that maximizes the coherence factor, a measure of how well echoes add up in phase. It needs no inversion. Any gain we claim must exceed it.
The strongest existing correction methods, applied independently per visit, are iterative CUTE as described by Zengqiu et al. and the Simson et al. learned estimator. Both are trained or tuned on our simulated media, with the same hyperparameter search budget and compute as the joint method. Their published single-visit figures are a 10.3 ± 5.60 m/s mean absolute error for the learned estimator and up to 18.2% coherence-factor gain for iterative CUTE. Neither paper reports longitudinal or radiomic results: there is no reported result found on Lacuna for radiomic-delta stability under any sound-speed correction.
We also include a learned beamformer in the spirit of DeepBF. It improves CNR from 1.69 to 2.16 over delay-and-sum on in-vivo carotid data and implicitly absorbs some aberration. We retrain it on the same simulations so that it sees the same path variation.
The key ablation is the joint objective with $\lambda = 0$. This is identical code and compute with the temporal coupling removed, so it isolates whether coupling the two visits, rather than better estimation per visit, does the work.
Beating the strongest baseline means two things. First, the joint pipeline must lower the median $R_f$ across texture features with non-overlapping bootstrap confidence intervals relative to independent iterative CUTE and the learned estimator. Second, it must raise stable-versus-changed classification AUC on held-out media and on the phantom pairs. A gain that appears only in simulation counts as partial.
Why not just…?
Why not just keep only radiomic features that are stable on test-retest scans? This is the most serious threat. A reviewer could argue that dropping high-$R_f$ features gives a robust panel with no physics at all. We test it directly. We select features by $R_f$ on training media and report classification AUC on the surviving subset under naive beamforming.
If the surviving features still detect biological change as well as the corrected pipeline does, the correction step is unnecessary for this task, and we will say so. We expect it to fall short for one reason we can measure: the texture features most sensitive to aberration overlap with those most sensitive to scatterer-density change. Discarding them should discard signal. The experiment settles this either way.
Why not just harmonize features statistically across visits or scanners? Batch-correction methods remove an average shift per acquisition group. Aberration, however, is patient- and visit-specific, because it depends on each person's fat layer and probe placement. We include a per-site location-scale feature harmonization baseline. It should help with the global speed offset but not with changes in path thickness. The factorial design shows which.
Why not just train the downstream model with blur and deformation augmentation? Augmentation makes a classifier ignore blur. It cannot tell blur caused by a changed path from blur caused by a softened lesion, and Simson et al. point out that it discards the sound-speed values themselves, which are diagnostic. The DeepBF-style baseline trained on varying media is the strongest form of this alternative. If it matches the joint method on AUC while lacking an explicit speed-change map, the contribution narrows to interpretability.
Why not just pick the best uniform speed per scan? If most of the artifact comes from a global offset, autofocus captures it cheaply. Our simulations separate global offsets from spatially structured path changes, so we can report where autofocus suffices. If it suffices everywhere, the core claim fails.
Decisive Checks
The first decisive check is whether, under naive beamforming, a substantial fraction of texture features show $R_f$ near 1. If they do not, the problem is minor and the program stops at a negative-result report. That report would still be useful, since no such measurement exists in the file.
The second is whether independent per-visit correction leaves a residual artifact that the joint objective removes. This is shown by the $\lambda = 0$ ablation on both simulated pairs and phantom layer swaps.
The third is whether FRD computed on radiomic deltas separates stable from changed pairs in ultrasound. Its CT/MRI out-of-distribution AUC of 0.94 is a reference, not a target, and it should improve after joint correction. Testing this also delivers the ultrasound and temporal extension of FRD that its authors left open.
Depth is a stated risk. Beyond 30 mm, the learned estimator's known degradation will likely limit every corrected pipeline, so all results are stratified by lesion depth.
References
- Iterative Pulse–Echo Tomography for Ultrasonic Image Correction
- Investigating pulse-echo sound speed estimation in breast ultrasound with deep learning
- Probing the speckle to estimate the effective speed of sound, a first step towards quantitative ultrasound imaging
- Fréchet radiomic distance (FRD): A versatile metric for comparing medical imaging datasets
- Hybrid Diffusion Model for Breast Ultrasound Image Augmentation
- DeltaNet: Conditional Medical Report Generation for COVID-19 Diagnosis
- CoCa-CXR: Contrastive Captioners Learn Strong Temporal Structures for Chest X-Ray Vision-Language Understanding
- OADAT: Experimental and Synthetic Clinical Optoacoustic Data for Standardized Image Processing
- Deep Learning-Based Universal Beamformer for Ultrasound Imaging