Clinical grading systems for ultrasound lesions, such as the five-level thyroid risk scale TI-RADS, are built from separate visual criteria that live in different places. Composition and echogenicity (how bright the tissue looks) are properties of the lesion's interior. Margin irregularity and extension beyond the gland are properties of the thin band where lesion meets surrounding tissue. Current low-data models make one of two choices. They feed the whole crop to a single pooled feature vector, which mixes interior and boundary signal with background clutter. Or they mask strictly inside the lesion, which removes clutter but also removes the boundary. We propose a third option. A cheap lesion outline is converted into a set of concentric soft rings: the core, the interface band, and the immediate surroundings. Image features are averaged separately within each ring, and each grading criterion learns which rings to read from. The claim is that this spatial division of labor is a strong enough prior to improve fine-grained grading when expert labels are scarce. It should also hold up when the outline itself comes from crude, nearly free supervision rather than expert pixel masks.
This matters because the hardest clinical calls sit at intermediate grades, such as TI-RADS 3 versus 4, and those calls turn mostly on margins. These classes are also the ones with the fewest and noisiest labels. A method that sends boundary evidence to the margin judgment explicitly, without demanding expert segmentation, targets exactly the part of the grading scale where small-sample models collapse.
Core Claim
Two existing results disagree about masking. The Role of Pleura and Adipose in Lung Ultrasound AI masks out superficial fat and muscle in lung ultrasound and raises four-class severity AUC from 0.6553 to 0.7742, because unmasked models learn body-habitus shortcuts. Lesion-Centric Latent Phenotypes from Segmentation Encoders for Breast Ultrasound Interpretability pools segmentation-encoder features strictly inside predicted breast lesion masks. It then needs a separate neuro-symbolic arbitration step to fix margin errors on discordant cases, where margin accuracy rises from 59.1% to 90.9%. We read the second result as evidence that interior-only pooling throws away the boundary signal and that the arbitration step is a patch for that loss.
Our hypothesis is that features pooled over distance rings, with each criterion choosing its rings, recover the boundary signal inside the network. This should remove the shortcut problem without needing post-hoc logic. The gain should concentrate in the margin criterion and in the intermediate risk levels, and it should grow as the labeled fraction shrinks.
Mechanism
Let $f(x)$ be the backbone feature at location $x$. Let $d(x)$ be the signed distance from $x$ to the predicted lesion boundary, divided by the lesion's equivalent radius so that rings scale with lesion size. Ring $k$ lies between normalized radii $r_k$ and $r_{k+1}$, with soft edges of width $\tau$. Its weight map and pooled feature are
$$w_k(x)=\sigma\!\left(\tfrac{d(x)-r_k}{\tau}\right)-\sigma\!\left(\tfrac{d(x)-r_{k+1}}{\tau}\right),\qquad z_k=\frac{\sum_x w_k(x)\,f(x)}{\sum_x w_k(x)},$$
where $\sigma$ is the logistic function. Each TI-RADS criterion $a$ reads its own mixture $\sum_k \alpha_{a,k} z_k$, where $\alpha_{a,\cdot}$ is a learned softmax over the rings. The final risk level is computed from the criterion representations.
The soft edges and the size normalization are what should make the method tolerant of imprecise outlines. A boundary misplaced by a few pixels shifts weight between neighboring rings rather than deleting information. The learned ring weights are also a built-in interpretability check: if the method works for the stated reason, the margin criterion should put its weight on the interface ring.
Research Program
The primary testbed is the STN thyroid dataset used by CMCNet: Aligning Ultrasound Image Embeddings with Textual TI-RADS Representations for Fine-Grained Thyroid Classification. It contains 600 nodules with paired views, labels for the five TI-RADS criterion categories, and manual bounding boxes. CMCNet supplies most of our training recipe. It aligns image features from an ultrasound foundation encoder (USFM, or BiomedCLIP) to text templates of the criterion scores using its center-margin contrastive loss, then discards the text encoder at inference. We keep that recipe and the backbone, and replace only the pooling: CMCNet pools the box crop globally, and we pool over rings.
The outline is the place where the expert-annotation bottleneck bites, since STN has boxes and not masks. Pixel masks would bring back the cost we want to avoid. We therefore study three outline sources of increasing cost. The first is an ellipse inscribed in the existing box, which is free. The second is a lesion segmentor trained on the public breast ultrasound mask sets BUSI and BUS-UCLM and applied to thyroid without adaptation. The third is a segmentor trained with dense local contrastive pretraining on unlabeled STN frames plus masks drawn for 5% of nodules, following the recipe of IDEAL. IDEAL reaches 0.738 Dice with 1.25% labels on cardiac MRI, but has never been used to feed a downstream grader.
Two prior results constrain the design. Optimizing Synthetic Data for Enhanced Pancreatic Tumor Segmentation shows that elastic boundary noise collapses segmentation Dice from 48.48% to 32.49%, which warns that boundary-dependent methods can be brittle. We test that brittleness directly instead of assuming it away. Scaling Supervision for Free finds that auxiliary anatomical supervision helps only when the structure is coupled to the disease. A lesion outline is the most strongly coupled structure available here, which is why we use it.
A secondary testbed checks whether ring pooling can replace the arbitration step of the lesion-centric breast model. We train on BUSI and BUS-UCLM and test externally on BUS-BRA. The task is margin and malignancy, plus multi-tier BI-RADS if those labels can be obtained for BUS-BRA; the research file does not establish that they exist.
Baselines
The naive reference point is the whole ultrasound frame, uncropped, with global average pooling over the same USFM backbone. It is trained with the same contrastive loss, splits, and number of tuning trials as every other arm. The lung ultrasound study predicts this will be the worst arm, and confirming that keeps the comparison honest.
The strongest existing method is CMCNet itself with its USFM backbone. It reports 75.17% balanced accuracy and 91.78% weighted F1 on five-class TI-RADS on STN (CMCNet). We rerun it on our splits, with the same encoder initialization, the same optimizer budget, and an equal number of hyperparameter trials. We report our rerun alongside the published figure. Beating it means a higher balanced accuracy than our rerun by more than the spread across five seeds, together with a gain in margin-criterion accuracy. A gain in weighted F1 alone does not count, because that metric is dominated by the common classes.
The interior-only baseline is mask-weighted pooling restricted to the lesion. This is the mechanism of the lesion-centric breast model and of the mask-weighted CT triage network. It isolates whether including the boundary matters at all.
Two ablations isolate the key ingredient. In the first, ring features are concatenated and shared by all criteria, so the rings are present but there is no routing. In the second, the rings have hard edges ($\tau \to 0$), so the routing is kept but the soft edges are removed. Every arm uses the same outline source within a given comparison.
Why not just…?
Why not just crop a larger box and let attention pooling find the boundary? This is the most serious threat. A dilated crop with an attention-pooling head can in principle learn ring-like weighting on its own. Our argument is about sample efficiency, not expressiveness: with 600 nodules, and far fewer in the intermediate risk levels, the structural prior should matter more than flexibility. We therefore run a dilated crop with learned attention pooling at 10%, 25%, 50%, and 100% of the training labels. If it matches ring pooling at every fraction, the claim is false and we will report it as such. If our hypothesis is right, the gap should open at the low fractions and close at full data.
Why not just add the mask as an extra input channel? A mask channel gives the network the outline but does not separate interior from boundary evidence in the pooled feature. It also forces the backbone input to change, which breaks the foundation-encoder initialization. We include it as an arm with matched tuning. Semixup offers indirect support: in knee osteoarthritis grading, Semixup needed localized pooling to preserve small structures that whole-map averaging washes out, which suggests that where features are pooled matters, not only what the input contains.
Why not just keep interior pooling and the neuro-symbolic gate? The gate works on discordant breast cases, but it is a hand-designed rule layer bolted on after the network. It was evaluated mainly in a binary malignancy setting. On the breast testbed we compare ring pooling without the gate against interior pooling with a reimplemented gate. If the gate still wins on margin accuracy, the rings have not absorbed its role, and we will say so.
Proposed Experiments
The main experiment crosses six pooling arms (whole frame, CMCNet box crop, interior-only, dilated crop with attention, mask channel, rings) with three outline sources and four label fractions on STN. Each configuration is repeated over five seeds with patient-level splits. The metrics are five-class balanced accuracy, per-class recall for TR3 and TR4, per-criterion accuracy for the five TI-RADS categories, and expected calibration error.
The robustness experiment applies elastic boundary noise of increasing strength to the predicted outlines, using the same noise protocol as the pancreatic tumor study. It compares how fast balanced accuracy degrades for hard-edged interior pooling, hard rings, and soft rings. The interpretability check records the learned ring weights for each criterion, with the prediction that margin weight concentrates on the interface ring and echogenicity weight on the core.
Three outcomes would decide the question. The first is ring pooling beating both the CMCNet rerun and the dilated-crop attention arm at the 10% and 25% fractions. The second is its gain appearing mainly in margin accuracy and TR3/TR4 recall. The third is the free inscribed-ellipse outline retaining most of the gain obtained with the trained segmentor. If the second outcome fails, rings help for reasons other than the boundary, and the core claim is wrong even if accuracy rises. If the third fails, the method does not escape the annotation bottleneck, and its value narrows to settings where masks already exist.
References
- The Role of Pleura and Adipose in Lung Ultrasound AI
- Lesion-Centric Latent Phenotypes from Segmentation Encoders for Breast Ultrasound Interpretability
- CMCNet: Aligning Ultrasound Image Embeddings with Textual TI-RADS Representations for Fine-Grained Thyroid Classification
- Ideal: Improved Dense Local Contrastive Learning For Semi-Supervised Medical Image Segmentation
- Optimizing Synthetic Data for Enhanced Pancreatic Tumor Segmentation
- Scaling Supervision for Free: Leveraging Universal Segmentation Models for Enhanced Medical Image Diagnosis
- Computed tomography–based COVID–19 triage through a deep neural network using mask–weighted global average pooling
- Semixup: In- and Out-of-Manifold Regularization for Deep Semi-Supervised Knee Osteoarthritis Severity Grading From Plain Radiographs