Whether a pancreatic tumor can be removed depends mostly on one geometric quantity: how far around a major abdominal vessel, such as the superior mesenteric artery or vein, the tumor wraps. That measure is called the contact angle, and roughly 180° is the clinical dividing line. Patients with borderline tumors often receive months of chemotherapy before anyone decides on surgery. At restaging, the question is a change question: did the tumor pull away from the vessel? Current automated systems answer a different question. They measure contact on one scan at a time. To use them across time, the scans must first be aligned, and alignment is least reliable exactly where it matters, at a vessel interface only a few voxels wide.
We propose to measure tumor–vessel contact in a coordinate system defined by the vessel itself. For each target vessel, the surrounding CT is resampled into a straightened tube: position along the vessel's centerline on one axis, angle around it on another, and radial distance on the third. Two scans of the same patient taken months apart share this coordinate system wherever the same vessel can be traced. This holds without whole-abdomen registration, because the vessel provides the alignment. The tube is also small enough to process in full 3D at native resolution. That directly addresses the context lost when 3D scans are cropped, downsampled or sliced. A paired-scan model then predicts the change in contact along the vessel, using the baseline scan as an explicit prior for reading the follow-up. The central claim is that predicting contact change in vessel-anchored coordinates gives more accurate restaging, and better prediction of whether surgery can achieve a clean margin, than measuring each scan independently and subtracting the results.
Core Claim and What Is New
Write $C_t(s,\theta)\in[0,1]$ for the predicted probability that tumor touches the vessel wall at centerline position $s$ and angle $\theta$ at scan time $t$. The contact arc at each position is
$$A_t(s)=\int_0^{2\pi} C_t(s,\theta)\,d\theta ,$$
and the surgical quantity is the maximum of $A_t(s)$ over $s$, compared against 180°. The model is trained to output $C_{t_1}$ and the change $C_{t_1}-C_{t_0}$ jointly from the tube volumes of both scans. The loss is applied on the angular contact map rather than on voxel intensities.
Three pieces of prior work supply the components, and none of them makes this measurement.
Segmentation-based Assessment of Tumor-Vessel Involvement segments tumor, veins and arteries with 3D U-Net variants. It adds an overlap loss at the tumor–vessel interface and computes contact angles with uncertainty bands, but only on single scans inside a cropped region. We reuse its interface-focused supervision and its contact-angle readout, and we change the representation and the input to a pair of scans.
CRONOS forecasts future 3D volumes in continuous time. It reports only intensity metrics, and it notes itself that such metrics do not guarantee validity for downstream tasks. This motivates our choice to supervise geometry rather than voxels.
DeepPrognosis predicts clean versus involved margins (R0 vs R1) from contrast phases a minute apart. That timescale is not the months-long therapy interval we target.
Two lines of work show that explicit alignment beats subtraction when reasoning about change: difference graphs over anatomical regions and regional cross-attention in CoCa-CXR. Both use 2D chest radiographs and anatomical regions defined by atlases. Our alignment unit is the vessel centerline, which is the structure the surgical decision is about.
The proposal also takes a position on a disagreement in the literature. Multimodal Assessment of Pancreatic Cancer Resectability reaches an AUC of 0.86 while using segmentation only as auxiliary training supervision. Viviers et al. instead argue that explicit interface geometry is necessary. We test this disagreement directly in the longitudinal setting, as an ablation rather than an assumption.
Research Program
Building the tubes. The first component extracts centerlines for the superior mesenteric artery and vein, the portal vein and the celiac trunk, from a vessel segmenter trained on each scan. This step inherits the broken-branch problem: ViG3D-UNet documents trunk fragmentation and branch omission in standard 3D networks. The target vessels are large trunks, not distal branches, so we expect tracing to succeed in most cases, and we will report how often it fails. Centerlines from the two time points are matched by curve alignment anchored at branch points, such as the origin of the superior mesenteric artery from the aorta. This yields a shared $s$ axis even when the tumor deforms the vessel.
The paired model. The second component is a small 3D network that takes both tube volumes. It is periodic in angle, and cross-attention runs along $s$ between the two time points. Training is in two stages. The network first learns $C_t$ on single scans, with interface supervision in the style of Viviers et al. It is then fine-tuned on pairs to predict the change map.
Calibrated readout. The third component is a calibrated readout: the maximum contact arc with an uncertainty interval, a predicted crossing of the 180° threshold, and a margin-status probability. An optional head adds serial clinical variables such as tumor-marker levels.
Data. This is the main risk. The file lists no public longitudinal pancreatic dataset with contact labels. The Viviers dataset has 195 scans from 99 patients, and we will check whether its multiple scans per patient include pre- and post-therapy pairs or are only contrast phases. In either case, the decisive evaluation needs an institutional cohort of borderline-resectable patients with baseline CT, post-therapy CT, expert contact annotations at both times, and surgical pathology margin status. The static components can be developed on public data in the meantime: tube extraction and vessel continuity on ImageCAS and 3D-ircadb-01, and single-scan contact accuracy on the Viviers data. We will also build a controlled synthetic test, in which tumors in static scans are shrunk or grown by known amounts and the scan is re-rendered. The ground-truth contact change is then exact.
Baselines
Do nothing. The naive reference copies the baseline scan's contact measurement forward, assuming no change. CRONOS shows that this kind of baseline is hard to beat for slow change; on Lumiere it reports 7.55 NRMSE against 8.38 for the copy-forward baseline. We run it on the same pairs. Beating it means higher accuracy on the 180°-crossing outcome, specifically among patients whose category actually changed. On stable patients, the copy-forward baseline is close to correct by construction.
Independent measurement and subtraction. This is the strongest competitor. The Viviers pipeline, comprising nnU-Net or a probabilistic 3D U-Net with overlap loss, is run separately on each scan, and the two angles are subtracted. On single scans it reports venous contact-angle correlation of R² = 0.60, with 88.2% sensitivity and 85.7% specificity for detecting involvement. No reported result was found on Lumiere for longitudinal contact-change accuracy of any method. We retrain it on exactly our training scans, with the same augmentation and the same hyperparameter search budget as our model. Beating it means lower mean absolute error in contact change and a significantly better crossing classification on held-out pairs. We also require that our model's single-scan contact R² does not fall below that pipeline's on the same scans, so the gain is not bought by losing per-scan accuracy.
Direct prediction from the post-therapy scan. This baseline is the Swin-UNETR model with clinical variables and auxiliary segmentation from Multimodal Assessment of Pancreatic Cancer Resectability, which reports AUC 0.86 internally and externally. It is trained on the post-therapy scan, alone or with the baseline scan stacked as extra input channels. For margin prediction against pathology, the reference is DeepPrognosis, which reports balanced accuracy 0.736 ± 0.141 for R0 vs R1 on its own cohort. That figure is not directly comparable, so we retrain both methods on our cohort.
Forecasting then measuring. A CRONOS-style generative model predicts the follow-up volume, and the Viviers measurement is applied to the generated volume. This tests whether generating voxels preserves interface geometry.
Ablations. The key ablation isolates the coordinate system: the identical paired network run on deformably registered Cartesian crops of matched size instead of tubes. A second ablation removes explicit contact supervision, so that only a margin label trains the model. This tests the Viviers-versus-Basel disagreement.
Why not just…?
Why not just run nnU-Net on each scan and subtract the angles? This is the real threat, and we treat it as the main comparison rather than a straw man. Its weaknesses are known for static scans. GATSegDiff reports a connectivity score of 2.36 for nnU-Net against 1.09 for a graph-guided model on 3D-ircadb-01, where 1.0 is perfect. Viviers et al. report tumor Dice of only 0.66. When two noisy angle estimates are subtracted, their errors add. If the paired tube model does not beat subtraction on patients whose category changed, the conclusion is that the tube representation is unnecessary. We would then report that and recommend the simpler pipeline.
Why not just predict "no change"? For stable patients it is nearly right. It is wrong by definition for exactly the patients whose surgical decision changes. We therefore report results separately for stable and category-changing pairs, so this baseline cannot look strong simply because most pairs are stable.
Why not just register the scans and subtract? Subtraction of aligned images ignores structure; difference graphs and CoCa-CXR both find that explicit regional alignment is needed to track change. Abdominal registration across months of tumor shrinkage is also least trustworthy at the vessel wall. The Cartesian-crop ablation measures this directly. If registered crops match tube coordinates, the tube idea adds nothing beyond the paired training.
Why not just use a 2D foundation model or slices packed into color channels? VLMs perform near chance on relative position in medical slices (Your other Left!). Packing slices into channels, as in HounsWorld, recovers only adjacent-slice context. A contact arc depends on the full ring of tissue around the vessel, which these representations do not hold together, so we do not expect this alternative to be competitive. We include it only on the synthetic test, as a cheap confirmation.
Decisive Checks
The first decisive check is the synthetic test. On scans with known induced contact change, the tube model should recover the change with lower error than independent measurement and subtraction, across tumor-shrinkage magnitudes from small to large. If it cannot do this with exact labels, the clinical study is not worth running.
The second check is on the clinical cohort: the error in contact change and the accuracy of 180°-crossing detection among category-changing patients, with patient-level bootstrap confidence intervals. The third is the AUC for pathology margin status. Here, our explicit-geometry model and its implicit-supervision ablation together settle whether explicit interface geometry matters once time is involved.
The tube representation must also not break vessel tracing. We will report clDice and connectivity of the traced trunks on 3D-ircadb-01 and ImageCAS against nnU-Net. For reference, ViG3D-UNet reports HD95 15.21 versus 17.60 for nnU-Net on ImageCAS. We will also report the fraction of patients for whom centerline matching fails, because the method cannot be applied to those patients.
References
- Segmentation-based Assessment of Tumor-Vessel Involvement for Surgical Resectability Prediction of Pancreatic Ductal Adenocarcinoma
- CRONOS: Continuous time reconstruction for 4D medical longitudinal series
- DeepPrognosis: Preoperative prediction of pancreatic cancer survival and surgical margin via comprehensive understanding of dynamic contrast-enhanced CT imaging and tumor-vascular contact parsing
- Expert Knowledge-Aware Image Difference Graph Representation Learning for Difference-Aware Medical Visual Question Answering
- CoCa-CXR: Contrastive Captioners Learn Strong Temporal Structures for Chest X-Ray Vision-Language Understanding
- Multimodal Assessment of Pancreatic Cancer Resectability Using Deep Learning
- ViG3D-UNet: Volumetric Vascular Connectivity-Aware Segmentation via 3D Vision Graph Representation
- ImageCAS: A large-scale dataset and benchmark for coronary artery segmentation based on computed tomography angiography images
- Continuous and complete liver vessel segmentation with graph-attention guided diffusion
- Your other Left! Vision-Language Models Fail to Identify Relative Positions in Medical Images
- HounsWorld: A Multimodal World Model for Hidden Patient-State Readout, Reconstruction, and Simulation