In pancreatic cancer, the amount of scar-like connective tissue (stroma) around tumor cells does not relate to survival the way most imaging models assume. A large digital-pathology study of 800 resected tumors found that patients with an intermediate stroma fraction lived about twice as long as patients with either very little or very much stroma. Survival was longest when intermediate stroma came with many lymphocytes. Today this can only be measured on the slide, after surgery. Before surgery, when the decision between going straight to resection and giving chemotherapy first is actually made, clinicians have only a contrast-enhanced CT scan. That scan records how iodine contrast flows into the tumor and slowly leaks out of it.
This note proposes to learn the tumor's stroma composition from the CT scan's contrast timing, and specifically to recover the non-monotone low/intermediate/high stratification, not a single "more fibrosis is worse" score. The model is supervised by distributions. Each resected slide yields many local stroma fractions, and the model is trained so that the spread of its voxel-level predictions inside the tumor matches the spread of the slide measurements. No CT voxel ever has to be aligned to a microscope pixel. The physical rationale is that delayed contrast retention reflects how much extracellular space a tissue has, and stroma is mostly extracellular space. CT therefore plausibly carries a stroma signal, while the lymphocyte signal is much weaker. The proposal says so and treats lymphocytes as a secondary, possibly failing target.
Core Claim
The claim is that a three-way stroma class (low, intermediate, high), with the same cutoffs that separated survival in pathology, can be predicted from preoperative multiphase contrast-enhanced CT. The claim has two further parts. First, the CT-predicted classes should reproduce part of the survival separation the slides show. Second, two ingredients carry the signal: per-voxel contrast-timing features, and distribution-level supervision. Each part is testable against simpler competitors, and the most important of those competitors is the textbook delayed-phase extracellular-volume calculation, discussed below.
Where This Sits Relative to Prior Work
Three groups of work define the gap.
The first group motivates the target. Stroma and lymphocytes identified by deep learning are independent predictors for survival in pancreatic cancer segments H&E slides into tumor, stroma and lymphocytes with a U-Net (mean Dice 87.74%). It defines Stroma-in-Percentage (SIP) and Lymphocytes-in-Percentage (LIP). It finds the intermediate-stroma survival advantage: 24 versus 12 months median overall survival in the Aachen cohort, and 28–35 versus 8–15 months when combined with high LIP. That work never touches imaging. We reuse its segmentation pipeline and its cutoffs as our label generator.
The second group has tried to read the microenvironment from CT, but with targets that cannot express a non-monotone effect.
- The value of a dual-energy CT Iodine map radiomics model for the prediction of collagen fiber content in the ccRCC tumor microenvironment classifies kidney tumors as high or low collagen from hand-crafted iodine-map features. It reports validation AUC 0.722 (SVM) and found the venous phase more informative than the arterial phase (0.741 vs 0.698). We take that phase finding as prior evidence that late contrast retention tracks matrix density. Its binary label cannot tell intermediate stroma apart from dense stroma, and that distinction is exactly what matters here.
- Non-invasive tumor microenvironment evaluation and treatment response prediction in gastric cancer using deep learning radiomics fuses a deep network with radiomics into one global score (TME classification AUC 0.909–0.937 across 2,686 patients). It also does not model local composition or contrast timing explicitly.
- AI-based non-invasive profiling of the tumor immune microenvironment using longitudinal CT radiomics predicts immunotherapy response in lung cancer predicts PD-L1 status (AUC 0.85). It needs at least two scans weeks apart to measure growth. We instead use the minutes-scale timing within a single contrast-enhanced session, which is available before any treatment.
The third group supplies tools and constraints.
- PanDx: AI-Assisted Early Detection of Pancreatic Ductal Adenocarcinoma on Contrast-Enhanced CT detects pancreatic tumors on about 3,000 portal-venous scans from the PANORAMA challenge (AUROC 0.9263). We reuse its coarse-to-fine segmentation recipe to obtain tumor masks. It says nothing about tissue composition.
- A Novel Framework for Characterization of Tumor-Immune Spatial Relationships in Tumor Microenvironment shows that immune cells cluster non-randomly within 100 µm of tumor nests. That scale is two orders of magnitude below a CT voxel. This constrains the proposal: we should not expect CT to resolve immune niches, so lymphocyte prediction is reported but not claimed.
- CO-Optimal Transport and Sliced Optimal Transport Plans offer machinery for matching distributions across unaligned spaces. We deliberately do not need them. Our matched quantity is a one-dimensional stroma fraction, so the transport distance has a closed form. Choosing the simpler tool is part of the design.
Research Program
Paired cohort. The first component is a paired cohort, because no public dataset in the material pairs preoperative contrast-enhanced CT with resected pancreatic slides. We would assemble a retrospective cohort of resected pancreatic adenocarcinoma at participating centers. The cohort would include every patient with:
- a preoperative pancreatic-protocol CT with at least two contrast phases;
- archived H&E slides from the resection;
- survival follow-up.
Delayed-phase acquisition is not routine in every pancreatic protocol. We would therefore define a primary subset with a delayed phase and a larger secondary subset with only the pancreatic and portal-venous phases. Results are reported separately for each subset, so readers can see how much depends on the delayed phase. Slides are processed with the Tan et al. pipeline, including Macenko color normalization. This produces a SIP and LIP value for every tissue tile, plus the slide-level class under their published cutoffs. Where several blocks per tumor exist, all are used, which gives a crude handle on within-tumor spread.
Contrast-timing representation. The second component turns the scan into per-voxel contrast-timing features. The phases are rigidly and then deformably registered to one another; this is CT to CT, which is routine, unlike CT to histology. Each tumor voxel's attenuation is normalized to the aorta in the same phase. From the result we compute wash-in, wash-out and delayed-retention values. The delayed-retention channel is an extracellular-volume proxy. Writing $\Delta T$ and $\Delta B$ for the delayed-minus-precontrast attenuation change in tissue and in blood, and $h$ for hematocrit, the standard estimate is
$$\mathrm{ECV} = (1-h)\,\frac{\Delta T}{\Delta B}.$$
These channels plus the raw phases feed a 3D network. Its output is a predicted stroma fraction $\hat s_v \in [0,1]$ for every tumor voxel $v$.
Distribution-level training. The third component is the training signal. For patient $i$, let $\hat F_i$ be the empirical distribution of $\hat s_v$ over the tumor mask, and let $G_i$ be the empirical distribution of tile-level SIP values from that patient's slides. Both are one-dimensional, so the 1-Wasserstein distance is the area between their quantile functions:
$$\mathcal{L}_i = \int_0^1 \big|\hat F_i^{-1}(q) - G_i^{-1}(q)\big|\,dq.$$
We add a cross-entropy term on the three-way class, computed from the predicted mean. This loss asks the model to reproduce how heterogeneous the tumor is, not only its average.
The underlying hypothesis is that "intermediate" slides are of two kinds:
- genuinely mixed tumors, in which dense and sparse regions coexist;
- uniformly moderate tumors.
If those two kinds have different contrast-timing signatures, a model that predicts only the mean will blur exactly the class that matters. The slide is a biased sample of the volume. The loss is therefore applied as a soft target, and we test how robust the results are to it, as described below.
Baselines
Clinical baseline. The naive reference point is a clinical model using only information available before surgery: tumor diameter from the CT mask, age, and radiological resectability category. It is fit by multinomial logistic regression for the stroma class and by Cox regression for survival. Any imaging model has to beat it to matter. Tan et al. report that their slide metrics carried information beyond stage, but no reported result was found on Lacuna for a preoperative clinical-only model predicting SIP class.
Single-formula ECV baseline. The simplest physics competitor is the mean tumor ECV from the formula above, with the class assigned by thresholds tuned on the training folds. This is the baseline we take most seriously. If one radiological number recovers the classes as well as our network does, the network is unnecessary. No reported result was found on Lacuna for ECV predicting pancreatic stroma class.
Hand-crafted radiomics. This follows The value of a dual-energy CT Iodine map radiomics model for the prediction of collagen fiber content in the ccRCC tumor microenvironment: per-phase texture and intensity features, then variance filtering, univariate selection and LASSO, then SVM and random forest. It is run on our masks and labels. That paper's validation AUC of 0.722 is on a different organ and a binary label, so it frames expectations but is not a target number.
Deep-learning radiomics. The strongest existing method is the hybrid of Non-invasive tumor microenvironment evaluation and treatment response prediction in gastric cancer using deep learning radiomics: a deep backbone with channel attention, fused with radiomics features and trained directly on the slide-level class. It reports TME classification AUC 0.909–0.937, again on a different task and cancer, so the number is context, not a bar we can claim to clear.
Fairness and the bar for success. All learned models receive the same tumor masks, the same phases, the same patient-level cross-validation folds, an equal hyperparameter search budget of 50 trials, and the same GPU allotment. "Beating" a baseline means higher macro-averaged one-vs-rest AUC for the three-way class on a held-out center, with a bootstrap confidence interval on the difference that excludes zero. The intermediate-class AUC is reported separately.
Ablations. Two ablations isolate the proposal's ingredients. The first removes contrast timing: the same network sees only the portal-venous phase. The second removes distribution supervision: the same network is trained on mean SIP regression plus class cross-entropy. As an upper reference, the slide-derived classes themselves are used for survival analysis on the same patients, which shows how much of the pathology signal CT recovers.
Why not just…?
Why not just biopsy the tumor and measure stroma on the sample? Preoperative pancreatic sampling is usually a fine-needle specimen from a deep, risky location. It is taken for diagnosis, and it samples a tiny focus of a tumor known to vary from region to region. Stroma and lymphocytes identified by deep learning are independent predictors for survival in pancreatic cancer derived its cutoffs from whole resection slides, not needles, so whether a biopsy reproduces the class is itself unknown. In patients who have both, we would compare biopsy-derived and CT-derived class agreement with the resection class. If biopsy agrees much better, the CT method has value only for patients who lack a usable biopsy, and we would say so.
Why not just regress mean stroma with a standard 3D network? This is the real threat, and it is the second ablation. Most cross-modal radiology–pathology work does exactly this, as in Non-invasive tumor microenvironment evaluation and treatment response prediction in gastric cancer using deep learning radiomics. Our argument that the mean blurs the intermediate class is a hypothesis, not a finding. If mean regression matches distribution matching on intermediate-class AUC and on survival separation, we will drop the distributional loss and report the contrast-timing channels as the only contribution.
Why not just compute ECV and threshold it? This could plausibly work, which is why it is a named baseline rather than a straw man. The non-monotone survival pattern raises a doubt. If intermediate and high stroma differ partly in cellular makeup (Tan et al. link intermediate stroma to inflammatory fibroblasts) rather than only in matrix amount, a single extracellular-space number may not separate them. That would leave room for wash-in and spatial-pattern features. If ECV alone wins, the honest conclusion is that a simple, already-validated radiology measurement predicts a strong pathology prognostic class. That would still be a useful result, and we would publish it as such.
Why not just register the CT to the slides and train voxel-by-voxel? Fixation shrinkage, sectioning, and unrecorded block orientation make this unreliable at scale. The thickness mismatch alone is about 1–5 mm voxels against 3–5 µm sections. Distribution-level supervision sidesteps registration entirely. General cross-space alignment methods such as CO-Optimal Transport are unnecessary because our matched variable is one-dimensional.
Proposed Experiments and Decisive Checks
Primary experiment. We use leave-one-center-out evaluation for the three-way stroma class, with macro AUC and intermediate-class AUC, comparing all baselines and ablations under the protocol above. The decisive outcome is the intermediate-class AUC of the full model against both mean-regression and ECV-threshold baselines. That class is the point of the study, and it is where monotone scores should fail.
Survival experiment. On held-out centers we stratify patients by CT-predicted class. We then test whether intermediate-predicted patients live longer than low- and high-predicted ones, using a log-rank test and a Cox model adjusted for the clinical covariates. We report how much of the slide-based hazard separation the CT classes retain. A CT model that classifies well but loses the survival pattern would fail the study's purpose.
Robustness to slide sampling. In tumors with multiple blocks, we test whether the predicted voxel distribution's spread correlates with between-block SIP variance. We also check whether results hold when training uses only one randomly chosen slide per patient. This probes whether the distributional loss captures within-tumor variation or merely fits slide noise.
Lymphocytes. LIP high/low prediction is reported with the same protocol. Given the 100 µm scale of immune clustering in A Novel Framework for Characterization of Tumor-Immune Spatial Relationships in Tumor Microenvironment, we expect near-chance performance. A null result here is informative: it would mark which parts of the microenvironment are and are not reachable from contrast CT.
Phase dependence. The full analysis is repeated on the subset lacking a delayed phase. This tells clinics whether adding a delayed acquisition is worth its extra radiation dose for this purpose.
References
- Stroma and lymphocytes identified by deep learning are independent predictors for survival in pancreatic cancer
- The value of a dual-energy CT Iodine map radiomics model for the prediction of collagen fiber content in the ccRCC tumor microenvironment
- Non-invasive tumor microenvironment evaluation and treatment response prediction in gastric cancer using deep learning radiomics
- AI-based non-invasive profiling of the tumor immune microenvironment using longitudinal CT radiomics predicts immunotherapy response in lung cancer
- PanDx: AI-Assisted Early Detection of Pancreatic Ductal Adenocarcinoma on Contrast-Enhanced CT
- A Novel Framework for Characterization of Tumor-Immune Spatial Relationships in Tumor Microenvironment
- CO-Optimal Transport
- Sliced Optimal Transport Plans