Many cancer imaging studies find that the tissue just outside a tumor predicts treatment response or survival better than the tumor itself. That tissue holds immune cells, remodeled stroma and new blood vessels. The standard way to measure it has three steps. An expert traces the tumor outline by hand. Software grows that outline outward by a fixed distance, such as 5 mm, to make a ring around the tumor. Texture statistics are then computed inside the ring. This recipe has two weaknesses. It needs dense expert tracing, which almost no routine clinical exam has. And it treats the tumor edge as a sharp line, although on CT and MRI the edge is often a low-contrast smear, which is also where different readers and different scanners disagree most.
We propose to compute peritumoral features without any traced outline. The only input is the two perpendicular diameter lines ("RECIST calipers") that radiologists already draw during routine tumor measurement. From those lines, a weakly supervised segmenter produces not one tumor outline but a calibrated spread of plausible outlines. Each peritumoral ring is then defined as an average over that spread. The resulting ring is soft where the edge is genuinely ambiguous and sharp where the edge is clear. The claim is that this uncertainty-averaged ring recovers most of the predictive value of rings built from expert outlines. It should also be more stable across readers and scanner manufacturers than hard rings grown from the same weak masks. Most importantly, it should beat a simple uniform blur of equal average width. If that last comparison fails, the method reduces to an existing recipe, and we will report that.
Core Claim
The paper that most directly motivates this work compared binary masks with "fuzzy" masks: Machine learning predictive performance evaluation of conventional and fuzzy radiomics in clinical cancer imaging cohorts. It blurred expert masks with a Gaussian whose width is set by the scanner's resolution. It reports AUC gains of +4.7% to +10.4% over binary masks, and an extended fuzzy margin reaching AUC 0.86 against 0.80 for binary on prostate PET/MRI. Its softness is uniform around the whole lesion and fixed by scanner physics. It also starts from a full expert delineation.
Our proposal differs on both counts. The softness is learned per voxel from how uncertain the edge actually is in that image. The starting point is two caliper lines rather than a traced outline. This changes which kind of error the ring absorbs. A uniform blur widens a clearly visible pleural or vessel interface just as much as a faint infiltrative margin. An uncertainty-averaged ring spends its softness only where readers would disagree.
A tension in the literature makes this worth testing. Towards automatical tumor segmentation in radiomics shows that human readers over-segment ambiguous edges as a safety margin, so peritumoral tissue leaks into "tumor" masks. Weakly supervised methods do the opposite. Learning to Segment Medical Images with Scribble-Supervision Alone removes uncertain edge pixels via Monte Carlo dropout and conditional-random-field refinement, and its remaining errors sit almost entirely on edges. So the zone that weak supervision discards as noise is the zone peritumoral radiomics wants to measure. Averaging over boundary uncertainty is a principled way to keep that zone instead of choosing one arbitrary edge.
Research Program
The first component is a caliper-supervised 3D lesion segmenter. It follows the slice-propagation approach of Accurate Weakly-Supervised Deep Lesion Segmentation using Large-Scale Clinical Annotations. That method turns RECIST lines into initial GrabCut masks, trains a network with a noise-tolerant loss, and propagates predictions to neighboring slices. It reports 91.5% Dice on the annotated key slice and 76.4% on full volumes. We reuse this pipeline rather than improve it. The one change is to train a small ensemble of five members with different seeds and pseudo-label perturbations, so that the segmenter yields a distribution of outlines $q(S)$ rather than a single mask. The segmenter is trained on DeepLesion's RECIST annotations.
The second component defines the rings. Let $S$ be one sampled outline, $d_S(x)$ the signed distance from voxel $x$ to that outline in millimetres, and $[r_k, r_{k+1})$ the $k$-th ring's distance interval (for example 0–5 mm, 5–10 mm). The weight of voxel $x$ in ring $k$ is
$$w_k(x) = \mathbb{E}_{S \sim q}\left[\mathbf{1}\{ r_k \le d_S(x) < r_{k+1} \}\right],$$
that is, the fraction of plausible outlines under which $x$ falls in ring $k$. Texture features are computed with these voxel weights, using weighted histograms and weighted co-occurrence counts, as the fuzzy-radiomics work does with its Gaussian masks. Where every ensemble member agrees on the edge, $w_k$ collapses to a hard ring. Where they disagree, the ring spreads out.
The third component calibrates the spread. The ring is only meaningful if the ensemble's disagreement matches real disagreement about where the edge is. LIDC-IDRI provides up to four independent radiologist outlines per nodule. We therefore calibrate one temperature parameter on the source scanners so that the predicted boundary band covers the reader outlines at the nominal rate. The temperature controls how widely $q$ spreads, and "coverage" means the fraction of reader-outline surface falling inside the band of outline positions to which $q$ gives at least 5% probability. An optional variant lets the temperature depend on voxel spacing, since slice thickness varies across sites.
This calibration differs from Improved post-hoc probability calibration for out-of-domain MRI segmentation. That method predicts per-pixel temperatures from test-time perturbation sensitivity plus a denoising-autoencoder shape prior. It lowers calibration error from 16.78% to 15.37% on a cardiac cine-to-contrast-enhanced MRI shift. Its authors' own framing limits it to regular organ shapes. Our target is coverage of real reader disagreement, which requires no shape prior and so applies to amorphous lesions.
Baselines
The naive reference comes in two forms, both using the same caliper-trained segmenter. The first uses intratumoral features only. The second applies fixed 5 mm concentric rings to the ensemble-mean binary mask, the geometry used in Pretreatment radiomic biomarker for immunotherapy responder prediction in stage IB–IV NSCLC. That study found peritumoral rings beat intratumoral features for immunotherapy response, with external AUC 0.75, though from manual outlines. Beating this reference means a statistically significant AUC gain (paired DeLong test) and higher cross-reader feature agreement.
The upper reference is the same fixed rings built from each individual LIDC reader's outline and from the reader consensus. This tells us how much predictive value the missing annotation costs. Success here means closing most of the gap, which we set as recovering at least 80% of the AUC margin that expert rings hold over intratumoral-only features.
The strongest existing competitor for the soft-ring idea is Gaussian fuzzy radiomics following Grahovac et al.. The Gaussian blur is applied to the weak ensemble-mean mask, with width derived from each scan's resolution. We also run a version tuned on the source folds. The reported figures there are +4.7% to +10.4% AUC over binary and 0.86 AUC on prostate PET/MRI. Those come from different cohorts, and no reported result was found on Lacuna for fuzzy peritumoral features on LIDC-IDRI.
The strongest competitor for scanner robustness is DAFIT from Optimal Statistical Incorporation of Independent Feature Stability Information into Radiomics Studies. DAFIT injects modeled reader and scanner noise into training features. It showed that naive stability filtering costs over 5% AUC and that uncorrected models overstate real-world performance by nearly 20%. No reported absolute DAFIT AUC was found on Lacuna, so we re-run it. DAFIT operates on features, so it can be combined with any ring definition. We report every ring type with and without it.
A promptable-segmenter baseline feeds the RECIST-derived bounding box to a Segment Anything-style model, in the spirit of the adaptation in CaPro, and then builds hard rings from its output.
The key ablation replaces $w_k$ with an isotropic blur whose width equals the average ensemble spread for that same lesion. This separates spatially varying uncertainty from mere softness. A second ablation swaps the ensemble for Monte Carlo dropout.
All feature sets feed the same classifier: an L1-regularized logistic model, with an identical nested cross-validation tuning budget and identical PyRadiomics settings. All segmenters get the same training data and GPU budget.
Why not just…?
Why not just dilate the weak mask by 5 mm? This is the naive baseline, so the experiment answers the question directly. The literature predicts it will lose. Weak masks err mostly at edges (Can et al.), and a hard dilation carries that error unchanged into every ring. The fuzzy-radiomics comparison of extended binary against extended fuzzy masks found the binary version more redundant and less predictive. Our "boundary jitter" check below measures how far the hard-ring AUC moves when the edge shifts by one or two voxels.
Why not just blur the weak mask with the scanner's point-spread function? This is the real threat. It is cheap, well founded, and already shown to help. If the matched-width isotropic ablation performs as well as uncertainty-averaged rings, the honest conclusion is that softness matters and its spatial pattern does not. The paper would then become a demonstration that caliper-only fuzzy peritumoral radiomics works. We expect the difference to appear mainly on lesions next to sharp interfaces such as pleura, vessels or fissures. There a uniform blur leaks the neighboring structure into the ring and an uncertainty-weighted ring does not. We will therefore report results stratified by whether a lesion touches such an interface.
Why not just filter out features that are unstable across readers and scanners? The DAFIT study shows on this very dataset that filtering discards discriminative features, costing over 5% AUC. We therefore compare against DAFIT itself rather than against filtering. Uncertainty averaging reduces instability at its source, the ring geometry, while DAFIT compensates after features are computed. Because the two are complementary, the relevant question is whether our rings improve on DAFIT with hard rings, and we test that combination.
Why not just prompt a foundation segmenter with the calipers? A promptable model returns one hard cutout with no calibrated edge spread, and it is not immune to domain shift. AMOS found over 20% Dice collapse on unseen MRI scanners across all architectures, with a plain 3D U-Net beating a transformer on CT. The prompted baseline tests whether its outlines are good enough that hard rings on them match our soft rings.
Proposed Experiments
The primary testbed is LIDC-IDRI. It has multiple reader outlines per nodule, several scanner manufacturers, and malignancy ratings, which we binarize as the endpoint while acknowledging that they are radiologist judgments rather than pathology. We simulate RECIST calipers from each reader's outline by taking the longest axial diameter and its longest perpendicular, then discard the outlines for everything except calibration and evaluation.
The first decisive check is reader-swap stability. We compute each ring's features from every reader's outline. We then measure the intraclass correlation between our caliper-derived features and the reader-mean features, and compare it with the agreement between individual readers. The second is leave-one-manufacturer-out AUC, reported with boundary-coverage on the held-out manufacturer so that calibration under shift is measured rather than assumed. The third is the boundary-jitter curve: AUC as a function of moving the ring origin inward or outward by one or two voxels. If peritumoral signal lives in the ambiguous band, as the safety-margin finding suggests, hard rings should show a steep curve and ours a flat one.
Segmentation quality is checked separately on DeepLesion's 1,000 manually outlined slices. There we must stay near the 90.3% Dice of Weakly Supervised Lesion Co-Segmentation on CT Scans and the WSSS key-slice result, confirming that the ensemble has not degraded the point estimate.
A secondary clinical test, contingent on data access, uses the public TCIA portions of the multi-site glioblastoma cohort of R-DepTH, which reports C-indices of 0.60–0.75 with manual outlines, and of the NSCLC immunotherapy cohort. For each we simulate calipers from the available outlines and ask whether caliper-only uncertainty-averaged rings keep the survival stratification that expert rings achieve.
The project fails its central hypothesis if the matched-width isotropic blur equals our rings on AUC, on reader-swap stability, and on held-out-manufacturer performance. It succeeds if caliper-only rings recover most of the expert-ring AUC and beat the blur on interface-adjacent lesions and on unseen manufacturers.
References
- Machine learning predictive performance evaluation of conventional and fuzzy radiomics in clinical cancer imaging cohorts
- Towards automatical tumor segmentation in radiomics: a comparative analysis of various methods and radiologists for both region extraction and downstream diagnosis
- Learning to Segment Medical Images with Scribble-Supervision Alone
- Accurate Weakly-Supervised Deep Lesion Segmentation Using Large-Scale Clinical Annotations: Slice-Propagated 3D Mask Generation from 2D RECIST
- Improved Post-hoc Probability Calibration for Out-of-Domain MRI Segmentation
- Pretreatment radiomic biomarker for immunotherapy responder prediction in stage IB–IV NSCLC (LCDigital-IO Study): a multicenter retrospective study
- Optimal Statistical Incorporation of Independent Feature Stability Information into Radiomics Studies
- CaPro: Curvilinear-aware Prompt Learning with Single Unlabeled Image for Cost-effective Curvilinear Structure Segmentation
- AMOS: A Large-Scale Abdominal Multi-Organ Benchmark for Versatile Medical Image Segmentation
- Weakly Supervised Lesion Co-Segmentation on Ct Scans
- Radiomic Deformation and Textural Heterogeneity (R-DepTH) Descriptor to Characterize Tumor Field Effect: Application to Survival Prediction in Glioblastoma