Blurry Tumor Edges as a 3D Prognostic Signal

When a pancreatic tumor grows into nearby fat and blood vessels, its edge on a CT scan is not a clean line. It is a smeared band several millimeters wide where tumor and healthy tissue mix. Segmentation methods treat that band as noise and try to erase it. Survival models mostly ignore it: they either look at a rough box around the tumor or at a hard mask cut out of the scan. We propose to measure the band directly. First, fit a 3D surface to the tumor. Then, at each point on that surface, sample a short line of image features running inward and outward along the surface normal. Record how wide the ambiguous zone is at that point and which major vessel the point faces. Finally, pool these points into a patient-level risk score for overall survival and for whether surgery will leave tumor behind at the resection margin.

The core bet is that prognosis depends on where and how the tumor edge is blurry, and a whole-box 3D network trained on about two hundred patients cannot reliably discover this by itself. If the bet is right, a surface-indexed, direction-aware description of the edge should add information beyond stage, size, and existing deep signatures. If it is wrong, a single scalar or a soft mask fed to a standard 3D network will match it, and the experiments are built to show which is the case.

Core Claim

Prior work splits the tumor edge problem in three incompatible ways. The first line cuts the edge off. DenseSharp+ in Probabilistic Radiomics pools deep features strictly inside the predicted lesion mask. It does model reader disagreement, but only for benign-versus-malignant labels. The second line keeps the surroundings without describing the edge. DeepPrognosis runs a 3D ResNet and a convolutional LSTM over a registered multi-phase bounding subvolume, and its signature stays independent of TNM stage, size, and CA19-9 (hazard ratio 1.435, p=0.002). The third line flattens the edge. Sequential Learning on Liver Tumor Boundary Semantics unwraps each 2D axial tumor contour into a 1D sequence in polar coordinates and shows that edge-point features predict capsular invasion and microvascular invasion better than standard segmentation. The operation is purely per-slice, and the unwrapping breaks on concave or multi-focal tumors. Meanwhile, boundary-focused segmenters such as PnPNet and the overlap losses of Combating Uncertainty with Novel Losses deliberately squeeze out edge uncertainty.

Our claim is that a 3D surface representation of the tumor–parenchyma interface predicts overall survival and resection-margin status better than these alternatives at realistic cohort sizes. The representation is built from per-vertex radial feature profiles, the local width of the ambiguous band, and position relative to named vessels. We further claim that the ambiguity band should be kept rather than removed. Sharpening the segmentation before feature extraction should hurt prognosis, not help it. The three parts that no single prior work combines are the 3D surface, the direction relative to vessels, and the ambiguity kept as signal.

Research Program

The pipeline has four stages. Each can be replaced by a simpler version for ablation.

The first stage is segmentation with a measurement of ambiguity. We train an ensemble of $K$ native 3D segmenters for pancreas, tumor, and the key peripancreatic vessels, starting from default nnU-Net configurations. We use native 3D rather than slice-wise models because Roth et al. showed that independent 2D pancreas predictions stack into fragmented, jagged 3D shapes. DB-MSMUNet reaches 87.59% Dice on MSD Pancreas from 2D slices but gives up cross-slice consistency. The ensemble's mean tumor probability $\bar p(x)$ at voxel $x$ is kept as a soft field and never thresholded away.

The second stage builds a surface and radial profiles. We extract a mesh at the $\bar p = 0.5$ level set and sample $N$ roughly uniform vertices $v$, each with outward normal $n_v$. At each vertex we record a radial profile: encoder features and multi-phase intensities sampled at offsets $t \in [-5, +10]$ mm along $n_v$. We also record the local ambiguity width

$$w(v) = \big|\{\, t : 0.2 < \bar p(x_v + t\, n_v) < 0.8 \,\}\big|,$$

which is the length along the normal over which the ensemble is unsure whether the voxel is tumor. The 0.2 and 0.8 cutoffs are tunable and will be varied in a sensitivity analysis. Each vertex also carries its distance and direction to the nearest segmented vessel, plus a label naming that vessel. Treating vertices as a point set avoids any global unwrapping, so concave and multi-focal tumors are handled without the failure mode of the polar approach.

The third stage aggregates. A small transformer over the vertex set pools the per-vertex tokens into a patient embedding. Each token combines the radial profile, $w(v)$, and vessel-relative coordinates. The model is trained with the Cox partial-likelihood loss for overall survival and binary cross-entropy for R0 versus R1 margin, the same two heads as DeepPrognosis. The segmenter stays frozen in the main model. We do this because co-training a segmenter through a survival loss on roughly 200 patients invites the overfitting seen in Glioma Prognosis, where survival Spearman correlation fell from 0.496 on validation to 0.148 on test. End-to-end co-training is tested only as a secondary ablation.

The fourth stage checks that the ambiguity measure is real. $w(v)$ is only useful if it reflects real image ambiguity rather than ensemble quirks. Before using it for prognosis, we test it on LIDC-IDRI, which provides up to four radiologist outlines per nodule. The test is whether per-vertex ensemble width correlates with per-vertex spread among the human outlines.

Data

The primary setting needs an institutional multi-phase CE-CT pancreatic adenocarcinoma cohort with overall survival and pathological margin status, comparable to the 205-patient cohort in DeepPrognosis. That cohort is not public, so we assume access to a similar local one, ideally split across two sites for an external test. Segmenters are pretrained on MSD Pancreas and the NIH Pancreas set, both described in The Medical Segmentation Decathlon and DB-MSMUNet.

BraTS 2018 survival (163 cases) provides a public second setting in which glioma edges are also infiltrative. We use it to check that the idea is not specific to the pancreas. LIDC-IDRI is used only for the ambiguity-validity check.

Baselines

The naive reference point is a Cox model on TNM stage, tumor size, and CA19-9, alongside a Cox model on automatically measured tumor volume. Any imaging model has to add to these, not replace them.

The strongest existing method is DeepPrognosis. It is a 3D ResNet18 with a contrast-phase ConvLSTM over a registered multi-phase subvolume around the tumor, trained jointly on survival and margin. It reports C-index 0.705 for overall survival and margin balanced accuracy 0.736 on its own cohort (DeepPrognosis). That figure comes from a different cohort and cannot be compared directly. We will therefore reimplement it on our patients and folds, using the same phase registration and subvolume size and the same number of hyperparameter trials as our model. Beating it means two things. The first is a higher C-index whose bootstrap 95% interval for the paired difference excludes zero across repeated cross-validation and on the external site. The second is that our signature stays significant in a multivariate Cox model that includes the DeepPrognosis signature itself.

The cheapest serious competitor is a soft-mask-channel network. This is the same 3D CNN given the ensemble tumor probability map and vessel masks as extra input channels. It sees exactly the ambiguity information our model uses, but has to learn the geometry on its own. It gets the same tuning budget.

Two shell-based baselines follow the peritumoral design of CT-Based Intratumoral and Peritumoral Radiomics Nomograms. There, a tumor-plus-10mm radiomics model reached external AUC 0.868 for spread through air spaces in lung cancer. We run both PyRadiomics features and pooled deep features in uniform 5 mm and 10 mm shells, fitted with LASSO-Cox.

A 2.5D boundary baseline applies the polar unwrapping of Sequential Learning on Liver Tumor Boundary Semantics to every axial slice and averages across slices.

Three ablations isolate our ingredients. The first removes $w(v)$ and uses hard-mask profiles only. The second replaces vessel-relative coordinates with orderless pooling. The third feeds the surface model a boundary sharpened by a PnPNet-style push–pull segmenter, which actively removes ambiguity. On BraTS the reference is the two-stage geometric pipeline of Glioma Prognosis, with test accuracy 46.8% and Spearman 0.148. No reported result was found on Lacuna for a stronger deep survival model on BraTS 2018.

Why not just…?

Why not just feed the soft tumor probability map to a 3D CNN? This is the most serious threat. It is cheap, and in principle it contains everything our model uses. Our argument concerns sample efficiency. With about 200 events-limited patients, a network must discover from scratch that the edge facing the superior mesenteric vessels matters more than the edge facing parenchyma. The surface representation hands it that structure. We cannot cite evidence that settles this; the paired comparison will. We will also train both models on nested subsets of 50, 100, and 200 patients. If the mask-channel CNN matches us at every size, we will conclude that the surface machinery is unnecessary and report that.

Why not just compute one number, the mean ambiguity width at vessel-facing surface, and add it to Cox? This is a real possibility, and a welcome one if true, because it would be a simple, auditable biomarker. Our first experiment tests exactly this scalar against margin status and survival before any learned aggregation is built. If the full model beats it by less than the bootstrap interval, the paper's contribution becomes the scalar.

Why not just use uniform peritumoral shells? Shells worked well for lung cancer spread through air spaces (CT-Based Intratumoral and Peritumoral Radiomics Nomograms). However, they average away direction. For pancreatic cancer, margin involvement is concentrated at the vessel interface, so a shell mixes the relevant 20% of the surface with the irrelevant 80%. The deep-feature shell baseline tests this, as does the ablation that removes vessel coordinates. If the shells tie with our model, direction is not carrying the signal.

Why not just use DeepPrognosis, whose subvolume already includes the surroundings? It does include them, and it is independently prognostic. The question is whether explicit edge description adds to it. We therefore test incremental value, reporting the C-index gain and the multivariate hazard ratio with the DeepPrognosis signature as a covariate. We do not test replacement.

Proposed Experiments

The first decisive check is cheap. On the pancreatic cohort, we test whether vessel-facing ambiguity width separates R1 from R0 patients and stratifies overall survival after adjusting for stage and size. In parallel, we test whether ensemble width tracks inter-rater spread on LIDC-IDRI. If both fail, the ambiguity half of the idea is dropped and only the direction-aware surface model is pursued.

The second check compares ambiguity kept against ambiguity removed. The identical surface model is fed either our soft ensemble boundary or a PnPNet-sharpened boundary. Our hypothesis predicts that sharpening lowers the C-index and margin AUC even though it raises segmentation surface accuracy. That would be direct evidence that edge-squeezing methods discard prognostic signal.

The third check is the main comparison. Every baseline is run on shared repeated cross-validation folds and a held-out external site. We report C-index, time-dependent AUC for overall survival, and balanced accuracy and sensitivity for margin, all with paired bootstrap intervals. The learning-curve study with the mask-channel CNN is included here.

The fourth check tests transfer. The same pipeline, with vessel coordinates replaced by distance to the ventricles and the cortex, is run on BraTS 2018 survival and compared against the geometric two-stage baseline. The aim is to see whether ambiguity-aware surface features close the gap between validation and test that sank that pipeline.

A secondary experiment unfreezes the segmenter and adds the survival loss. It asks whether prognostic supervision moves the surface in clinically meaningful places, which we check by comparing against pathology-reported margin sites where available, and whether it harms generalization.