A surgeon who removes a cancer "en bloc", meaning in one intact piece wrapped in healthy tissue, gets a much better result than one who removes it in fragments. In spinal giant cell tumors, local recurrence falls from 36.7% to 9.5% (A Radiologists' Guide to En Bloc Resection of Primary Tumors in the Spine). Whether a clean margin was achieved, called R0 (no tumor at the cut edge) versus R1 (microscopic tumor at the edge), is only known after surgery, once a pathologist has examined the specimen. A breached margin cannot be repaired afterward, and adjuvant radiation does not make up for an incomplete resection (The Oncological Mechanics of En Bloc Resection). For this reason the decision to operate now, or to give more therapy first, has to be made from preoperative imaging.
Current imaging models make that decision from one visit. Patients who receive therapy before surgery (neoadjuvant therapy) usually have several scans spread over months. On the final scan, residual tumor and treatment scar can look alike where they touch the major vessels. We propose that the trajectory across those scans carries margin information the last scan does not. The key signal is how the tumor–vessel contact surface changes over time: whether it shrinks, holds, or advances. We will test this in pancreatic cancer treated with neoadjuvant therapy. The primary test uses a deliberately simple interface-trajectory model. A continuous-time neural model is included only as an extension that must earn its place.
Core Claim
For patients who undergo resection after neoadjuvant therapy, we claim that geometric features of the tumor–vessel interface, tracked across multiple dated preoperative CT scans, predict pathological R1 margins better than the best single-visit model on the final scan. By single-visit model we include one that already exploits multi-phase contrast within that scan.
The closest existing work is DeepPrognosis. It predicts R0 versus R1 in resectable pancreatic ductal adenocarcinoma using a 3D ConvLSTM over three contrast phases acquired minutes apart. It reports balanced accuracy 0.736 and R1 sensitivity 0.813 on 205 patients. Its time axis is contrast wash-in and wash-out within one session. Ours is disease evolution across weeks to months. It learns an unstructured deep signature. We track an explicit, anatomically defined quantity that a surgeon already reasons about: vascular encasement, which Smith et al. treat as a planning deal-breaker.
The second anchor, the spine radiology guide, supplies the clinical target. It is qualitative and single-timepoint. It has no computational model and no follow-up over time.
Research Program
The first step is reliable per-visit anatomy. We will train an anatomy-aware segmenter on PanTS. That dataset has 36,390 CT scans labeling the tumor, the pancreas, and 24 surrounding structures. PanTS showed that supervising those surrounding structures improves tumor Dice by 10.3% and tumor boundary accuracy by 9.7% over a tumor-and-pancreas-only model. This is exactly the boundary precision that interface measurement needs.
At each visit we will then compute interface descriptors for the vessels that decide pancreatic resectability. Those vessels are the superior mesenteric artery and vein, the portal vein, the celiac axis, and the common hepatic artery. The descriptors are:
- the arc of circumferential contact in degrees;
- the length of contact along the vessel;
- the minimum fat-plane distance;
- the attenuation of the soft tissue at the interface.
The attenuation feature is our proxy for fibrotic versus viable tissue. This descriptor set mirrors the structured WBB-style reporting logic of the spine guide, carried over to vessels.
The trajectory representation treats each descriptor as an irregularly timed series. We encode the series by its last value, its slope per month, and its change since the first scan, all normalized by the actual time between scans. A small monotone gradient-boosted model or logistic model maps these features, together with standard covariates, to R1 probability.
The extension replaces the hand-built slope features with a continuous-time encoder in the spirit of CRONOS and Continuous-Time Deep Glioma Growth Models. Both handle irregular scan intervals. We apply the encoder to interface maps rather than whole volumes, which also avoids the GPU memory failures those papers report for full 4D volumes.
Data is the main practical obstacle. No dataset in our source material pairs multiple dated preoperative CTs with margin pathology. We will therefore assemble a retrospective institutional cohort of neoadjuvant-treated pancreatic cancer patients. Each patient must have at least two preoperative CTs and a recorded R0 or R1 result.
This cohort has a built-in selection bias of the kind Smith et al. warn about for spine surgery: margin labels exist only for patients who were operated on. We therefore state our claim as conditional on resection. We treat any conclusion about the unresected population as out of scope.
Baselines
Clinical reference point. The naive reference is the clinician's own standard: the final-scan radiology resectability category together with CA19-9. We fit it by logistic regression on the same training folds.
Strongest existing method. This is a reimplementation of the DeepPrognosis margin branch: a 3D-ResNet18 feeding a ConvLSTM across the contrast phases of the final preoperative scan, trained multi-task with survival. It is the strongest competitor because it is the only published preoperative margin predictor in our sources, and it already exploits a form of temporal contrast.
For reference, DeepPrognosis reports balanced accuracy 0.736 and R1 sensitivity 0.813 on its own cohort. That was an upfront-resection cohort, not a neoadjuvant one. Since no result has been reported on our cohort, we will retrain it ourselves.
Fairness conditions. Every model gets:
- the same patients and the same patient-level cross-validation folds;
- the same PanTS-pretrained encoder weights where an encoder is used;
- a matched budget of hyperparameter trials.
Beating the strongest baseline means an R1 AUC improvement whose confidence interval, from a paired bootstrap over patients, excludes zero, plus higher R1 sensitivity at matched specificity.
Ablations. We add a static geometry baseline: the same interface descriptors computed on the final scan only, fed to the same small model. We also run two ablations that isolate the ingredient we claim matters. The first shuffles each patient's scan dates. The second replaces the history with copies of the last scan. If performance survives either manipulation, the trajectory is not what is carrying the signal.
Why not just…?
Why not just use the last preoperative scan? This is the real threat. CRONOS found that simply repeating the last image beats ConvLSTM and SimVP on longitudinal forecasting. It was beaten only by a carefully formulated flow model, with NRMSE 7.55 versus 8.38 on the Lumiere glioma dataset.
Our static-geometry baseline and our duplicated-history ablation exist to settle this directly. If the final-scan geometry matches the trajectory model, we will report that longitudinal history adds nothing for margin prediction. We will then redirect the work toward a better last-scan scar-versus-tumor classifier.
Why not just stack the scans as input channels? Channel stacking assumes a fixed number of evenly spaced visits, which real neoadjuvant schedules do not provide. DeepPrognosis also found stacking inferior to sequential modeling, even with fixed phases. We still include a stacked-channel 3D CNN on the two most recent scans as a baseline, so this claim is tested on our data rather than assumed.
Why not just learn a deep signature end to end, as DeepPrognosis does? With a few hundred patients, an end-to-end longitudinal network risks memorizing patient-specific anatomy. That is precisely the "normalcy bias" reported in the longitudinal imaging literature. Explicit interface features give the model an anatomical inductive bias, and they are auditable by surgeons. The continuous-time extension is the test of whether end-to-end learning adds anything once that bias is supplied.
Why not just prompt a general vision-language model with the scan series? Medical Large Vision Language Models with Multi-Image Visual Ability reports that generalist and single-image medical models fail at comparative multi-image diagnosis. Such models also cannot ingest full 3D series within memory. We will not spend budget on this baseline beyond a small qualitative probe.
Proposed Experiments
Primary experiment. We use five-fold patient-level cross-validation on the institutional cohort. The outcomes are:
- R1 AUC;
- balanced accuracy;
- sensitivity at 80% specificity;
- calibration slope.
We compare the trajectory model against all baselines above.
Decisive check. We stratify by final-scan geometry, focusing on patients whose last scan shows ambiguous vessel contact, meaning contact without clear encasement. The hypothesis predicts the gain concentrates there. Scar and tumor are hardest to separate in that group, which is the post-treatment ambiguity that BioViL-T documents in chest imaging as snapshot blindness. A uniform or absent gain would weaken the scar-disambiguation explanation even if overall AUC improves.
Supporting checks. We verify segmentation quality on the PanTS out-of-distribution test split. PanTS itself reports detection AUC 0.959. We will not report a figure for our interface-specific accuracy until it is measured.
We also test on Lumiere whether interface-trajectory forecasting of future tumor-to-structure contact beats last-image persistence. This is a methodological sanity check that uses public data. It has no margin labels.
Finally, a survival analysis tests whether the predicted R1 probability stays prognostic after Cox adjustment for stage and CA19-9. This mirrors the independence test DeepPrognosis reports (hazard ratio 1.435).
References
- A Radiologists' Guide to En Bloc Resection of Primary Tumors in the Spine: What Does the Surgeon Want to Know?
- DeepPrognosis: Preoperative Prediction of Pancreatic Cancer Survival and Surgical Margin via Contrast-Enhanced CT Imaging
- PanTS: The Pancreatic Tumor Segmentation Dataset
- CRONOS: Continuous time reconstruction for 4D medical longitudinal series
- Continuous-Time Deep Glioma Growth Models
- Medical Large Vision Language Models with Multi-image Visual Ability