An ultrasound image does not have one resolution. The sound pulse loses its high frequencies as it travels deeper into tissue, and a fan-shaped probe spreads its scan lines farther apart with depth. The smallest blob the scanner can tell apart is therefore a few tenths of a millimetre near the skin and several times larger at the bottom of the image. Segmentation networks ignore this. They are trained to match a hand-drawn outline exactly, with the same penalty at every pixel, and they are scored with boundary metrics such as Hausdorff distance that treat a 2 mm miss near the probe the same as a 2 mm miss deep in the image, where 2 mm may be below what the scanner can resolve. The result is that training spends effort fitting annotator noise in places where the image carries no boundary information. Evaluation then rewards or punishes that noise.
We propose to read an estimate of the local resolution off each image and use it in two places. First, the hard training contour is replaced by a soft band whose width and direction match the local resolution cell. Second, boundary error is reported in units of that cell. The resolution estimate comes from the speckle grain size: speckle is the granular texture that tissue produces in ultrasound, and the size of its grains tracks the scanner's point-spread function, meaning the blur a single point would show. The estimate needs no probe metadata and no raw radio-frequency data. The claim to be tested is specific. Supervision that varies with resolution across the image should improve boundary accuracy on deep and fan-probe regions beyond what a uniform soft boundary or a depth input channel achieves, and the predicted boundary uncertainty should track where the physics says the boundary is unknowable.
Core Claim
The resolution–penetration trade-off sets a spatially varying floor on how precisely a boundary can be located. Low-contrast boundary ambiguity is, in large part, that floor showing up in the image. Current boundary-aware networks treat ambiguity as stationary across the image. FreqDINO splits a frozen DINOv3 encoder's features into wavelet sub-bands uniformly across the image and decodes a 1-pixel contour. BPAT-UNet and Niu et al.'s contour-loss network supervise auxiliary branches with Canny edges computed from the binary masks. In all three, a deep, blurred boundary receives the same sharp target as a shallow, crisp one.
Our proposal differs from them in what it changes. We do not add a new architecture. We change the target and the metric so that both follow the resolution map. The proposal is closest in spirit to Bounding Maps, which replaced binary boxes with linearly decaying targets on CT and raised Faster R-CNN sensitivity at 4 false positives per image from 82.43% to 86.28%. Their decay is fixed by object geometry, whereas ours is set by measured image physics and changes across the image.
Research Program
Estimating local resolution. In each sliding window of the B-mode image (the standard grayscale ultrasound display), we compute the normalized autocorrelation of the speckle. The width of its central peak along the depth axis gives $\sigma_{ax}(x)$, the axial resolution at pixel $x$. The width across the image gives $\sigma_{lat}(x)$, the lateral resolution. Strong reflectors and anechoic regions (areas with no echo) are masked out, and the field is smoothed into a resolution map.
Log compression and vendor smoothing distort speckle statistics, so the estimator is first validated on simulated phantoms where the true point-spread function is known. On real data we check that the estimate grows with depth, and that it grows faster laterally on convex-probe frames than on linear-probe frames of ERUS-10K. If the estimator fails these checks, the project falls back to an analytic depth model. That fallback is weaker, because it needs probe parameters the public datasets lack.
Building the soft boundary target. Let $d(x)$ be the signed distance from pixel $x$ to the annotated contour, and let $n(x) = (n_z, n_x)$ be the unit normal of the nearest contour point, with $n_z$ its depth component and $n_x$ its lateral component. The resolution across the boundary at that point is $\sigma_n^2 = n_z^2\sigma_{ax}^2 + n_x^2\sigma_{lat}^2$. The soft mask target is
$$t(x) = \Phi\!\left(\frac{d(x)}{\kappa\,\sigma_n(x)}\right),$$
where $\Phi$ is the standard normal cumulative distribution and $\kappa$ is a single global scale tuned on validation data. The anisotropy matters. A contour running parallel to the probe face is limited by axial resolution, while a contour running along the beam is limited by lateral resolution, which is usually worse.
The network is trained with soft cross-entropy against $t$. It also has a second output head that predicts a per-pixel boundary width $\hat\sigma$, regressed toward $\kappa\sigma_n$ near the contour. That head lets us test whether the network's uncertainty becomes physically calibrated.
Evaluation in resolution units. We report a resolution-normalized boundary F-score, in which a predicted boundary pixel counts as correct if it lies within $\sigma_n$ of the reference contour. We also report the 95th-percentile Hausdorff distance (HD95) divided by local $\sigma_n$. Every boundary metric is additionally stratified into depth terciles and by probe geometry.
Physics motivates this design. Gare et al.'s confidence-annotated lung study showed that expert disagreement concentrates at transition zones and that training on pixels with at least 60% consensus improved downstream clinical prediction, but it needed costly multi-expert confidence maps. If the speckle-derived map predicts where experts disagree, the method provides a cheap substitute for those annotations. Three prior results constrain the design. Wu et al.'s TOFD enhancement shows that sparse deconvolution and synthetic-aperture focusing can recover resolution only under linear, shift-invariant blur, which is why we do not try to sharpen the image. The 50% positive predictive value that Yu et al. report for 1–2 mm melanomas at 18 MHz is a reminder that some boundaries are genuinely below resolution and should be predicted as uncertain, not sharp. Finally, Gare et al. 2021 found that models exploit superficial high-contrast layers when deep boundaries are blurred. That finding warns that metrics pooled over the whole image can hide failure at depth, which is why every result is depth-stratified.
Baselines
The naive reference is a standard nnU-Net trained with Dice plus cross-entropy on hard masks. FreqDINO reports this model at 84.80% Dice and 46.63 mm Hausdorff distance on BUSI (Breast Ultrasound Images). We retrain it on our splits.
The strongest boundary-aware competitor on BUSI is FreqDINO itself, which reports 86.52% Dice and 39.63 mm Hausdorff distance on BUSI and 62.09% Dice zero-shot on TN3K. It is the strongest competitor because it already targets the separation of boundary from speckle in the frequency domain, which is the closest existing answer to our question. BUSSAM, a SAM adaptation with a parallel CNN branch, reports 89.95% Dice on BUSI. However, it uses box prompts and an evaluation protocol the file does not let us match, so we rerun it under our protocol rather than quote it as the target.
On TN3K, the reference boundary method is BPAT-UNet, which reports 83.64% Dice similarity coefficient and HD95 of 14.06 against U-Net's 80.79% and 21.51. On ERUS-10K it is ASTR, which reports 77.6% Dice against MedSAM's 75.7%.
All methods are retrained on identical splits with the same augmentation, the same epoch budget, and an equal hyperparameter-search budget. Our supervision is applied as a drop-in loss change to two backbones: nnU-Net, and FreqDINO's own decoder. This separates the effect of the target from the effect of the architecture.
Two ablations isolate the key ingredient. The first uses a uniform soft band with the same mean width as our resolution map, which tests whether any softening suffices. The second feeds the resolution map to the network as an extra input channel while keeping hard targets, which tests whether the information helps as input rather than as supervision.
We count the proposal as beating the baselines if it meets three conditions over five seeds on BUSI, TN3K and ERUS-10K. It must improve the resolution-normalized boundary F-score in the deepest tercile and on convex-probe ERUS frames significantly more than both ablations. It must not lose Dice overall. It must yield a predicted-width map $\hat\sigma$ that correlates with $\sigma_n$ on held-out data.
Why not just…?
Why not just give the network depth as an input channel and let it learn the rest? This is the most serious threat. A coordinate channel lets the network learn depth-dependent behaviour, and our resolution-as-input ablation is a stronger version of the same idea. Input conditioning, however, cannot change what the loss rewards: with hard targets the network is still pushed to commit to a sharp edge wherever the annotator drew one. If the input-channel ablation matches our result in the deep tercile, the supervision story is wrong and we will report that result.
Why not just smooth the labels uniformly or use a Bounding-Maps-style decay? Bounding Maps and the confidence-weighted losses of Gare et al. show that softening helps, so a fixed-width band is a real competitor. Our claim lives entirely in the spatial variation of the band. The uniform-band ablation is designed to show whether that variation matters, and the depth-stratified analysis shows where it matters: a uniform band should over-blur shallow edges and under-blur deep ones.
Why not just preprocess with gain compensation or despeckling and keep standard training? Gain compensation restores brightness lost with depth but does not restore resolution lost with depth. Despeckling removes the very texture our estimator reads. As Wu et al. show, recovering resolution after acquisition needs strong linear-blur and sparsity assumptions that tissue violates. We include a despeckled-input nnU-Net run to check this empirically.
Why not just fine-tune a foundation model? MedSAM trails ASTR on ERUS-10K, at 75.7% versus 77.6% Dice, and FreqDINO's motivation is that natural-image encoders do not distinguish acoustic noise from edges. Rather than argue against foundation models, we apply our loss on top of FreqDINO's frozen-encoder setup. Gains there would show that the two approaches are complementary.
Proposed Experiments
The first decisive check is the validity of the estimator. On simulated phantoms with known point-spread functions, the speckle-derived $\sigma_{ax}$ and $\sigma_{lat}$ should track the ground truth across depth. On ERUS-10K, lateral width should grow faster with depth on convex-array frames than on linear-array frames. A failure here stops the project before any training.
The second check is the main comparison on BUSI, TN3K and ERUS-10K described under Baselines. The emphasis is on the deepest tercile and on convex-probe frames, where the hypothesis predicts the gains concentrate. Flat gains across depth would argue against the physical explanation, even if overall Dice improved.
The third check is cross-dataset transfer. We train on BUSI and test zero-shot on TN3K, mirroring FreqDINO's protocol, where it reports 62.09% Dice. The prediction is that resolution-aware targets transfer better because speckle statistics are re-estimated on each new scanner rather than learned.
The fourth check, conditional on collaborator access to the confidence-annotated lung cohort, tests whether $\sigma_n$ and the predicted width $\hat\sigma$ correlate with per-pixel expert disagreement. A positive result would mean the method supplies, at no annotation cost, the uncertainty signal that Gare et al. had to collect by hand. No reported result on Lacuna was found for this correlation.
References
- Freqdino: Frequency-Guided Adaptation for Generalized Boundary-Aware Ultrasound Image Segmentation
- BPAT-UNet: Boundary preserving assembled transformer UNet for ultrasound thyroid nodule segmentation
- A Medical Image Segmentation Method based on Multi-scale Features and Contour Loss Constrain (S)
- Bounding Maps for Universal Lesion Detection
- Label Uncertainty for Ultrasound Segmentation
- Ultrasonic Time-of-Flight Diffraction Imaging Enhancement for Pipeline Girth Weld Testing via Time-Domain Sparse Deconvolution and Frequency-Domain Synthetic Aperture Focusing
- The utility of high-frequency 18 MHz ultrasonography for preoperative evaluation of acral melanoma thickness in Chinese patients
- The Role of Pleura and Adipose in Lung Ultrasound AI
- Ultrasound SAM Adapter: Adapting SAM for Breast Lesion Segmentation in Ultrasound Images
- Towards a Benchmark for Colorectal Cancer Segmentation in Endorectal Ultrasound Videos: Dataset and Model Development