Spatial Configuration of Pancreatic Cancer Is Associated with Disease Recurrence after Neoadjuvant Therapy and Curative-Intent Resection

Abstract

Purpose:

Pancreatic ductal adenocarcinoma (PDAC) frequently recurs after neoadjuvant therapy (NT) and curative-intent resection. Although major pathologic response predicts favorable outcomes, most patients achieve only minor response with heterogeneous recurrence risk. We asked whether the spatial organization of residual PDAC encodes clinically relevant biology beyond residual tumor burden.

Experimental Design:

In a retrospective cohort of 203 resected PDAC patients treated with NT and restricted to minor pathologic response, routine H&E whole-slide images were segmented into cancer and stroma using an AI-enabled pipeline. Spatial composition (patch density, edge density) and configuration (compactness/complexity, intermixing) were quantified and tested associations with disease-free survival (DFS) using multivariable models adjusted for standard clinicopathologic factors.

Results:

Shorter DFS was associated with a fragmented, interface-rich tumor-stroma ecology featuring higher edge density and diversity and reduced homotypic aggregation, independent of clinicopathologic variables. Two spatial risk models were independently prognostic: (i) cancer mean shape index plus stromal shape-index variability (adjusted HR 1.71; P=0.003) and (ii) mean stromal patch area plus edge density (adjusted HR 2.19; P=0.002). Both models stratified outcomes where pathologic response grading and residual cancer area did not. High-risk configurations were further associated with reduced intratumoral tumor-infiltrating lymphocyte (TIL) density and infiltration ratio, with relative TIL accumulation at the cancer periphery and within the stroma, consistent with an immune-excluded phenotype.

Conclusions:

Residual cancer–stroma topology quantified from standard H&E slides yields independent prognostic signals after NT in PDAC, provides a cellular immune correlate for spatial risk, and motivates prospective validation and spatially informed adjuvant strategies.

Introduction

Despite therapeutic advances, pancreatic ductal adenocarcinoma (PDAC) remains among the most lethal cancers, with high rates of recurrence despite extensive neoadjuvant therapy (NT) and curative-intent resection¹. While NT has improved treatment completion^2,3, margin-negative resection (R0) rates^4,5, nodal clearance⁶, pathologic response⁷, and has been associated in some studies with improved survival^4,6,8, definitive survival benefit remains under active investigation, and more than half of patients develop disease recurrence within two years⁹.

Pathologic treatment response correlates with outcome at the extremes: a major response is favorable^10–15, whereas most patients achieve only a minor response with heterogeneous courses thereafter⁹. Within this large majority with residual disease, current assessments emphasize how much tumor persists rather than how it is organized, relying on semi-quantitative or subjective scoring varying across observers^12,13. These limitations constrain risk stratification and do not consider spatial organization of disease.

Tumors are structured ecosystems¹⁶ in which cancer and stroma form patches and interfaces¹⁷ influencing nutrient flow^18–20, immune access^21–25, mechanical constraints^26–28, and routes for invasion^29,30. After therapy, selection can re-shape this ecology, potentially enriching spatial configurations that enable disease persistence³¹. Conventional pathologic treatment response assessments, capturing residual disease burden, may be enhanced through quantification of emergent topologies (e.g., fragmentation, interface density, homotypic clustering, or shape irregularity) that may encode clinically relevant biological information related to therapeutic resistance, disease persistence, and risk of relapse.

Here, focusing deliberately on the challenging and clinically prevalent cases of PDAC with minor pathologic responses after NT and resection, we ask whether residual cancer-stroma spatial organization in the resected tissue predicts disease-free survival (DFS) beyond standard clinicopathologic variables and gross disease extent. Using H&E whole-slide overlays to delineate cancer and stroma, we operationalize a computational pipeline (Cisneros et al., bioRxiv, 2025; DOI: 10.1101/2025.04.22.646608) to quantify composition (e.g., patch size/density, edge density) and configuration (e.g., patch complexity/compactness, spatial intermixing) at patch and landscape scales, and we develop multivariable spatial risk models. By reframing treated PDAC as a post-therapy landscape, rather than only focusing on disease burden, we aim to provide a quantitative, observer-independent description of the spatial tissue organization persisting after NT, furnish prognostic insights in the context of most cases with minor pathologic responses and generate testable hypotheses about the biology of residual disease.

Methods

Study Design, Study Population, and Data Collection

All studies involving human tissue or clinical data were conducted in accordance with the Declaration of Helsinki and were approved by the Mayo Clinic Institutional Review Board (IRB #22–007646). This retrospective minimal-risk study used archived residual biospecimens and clinical data and involved no subject interaction or intervention. The Mayo Clinic IRB approved a waiver of study-specific informed consent and a waiver of HIPAA authorization. Data and specimens were de-identified for analysis, and patients who had declined institutional research authorization were excluded.

All subjects had undergone surgical resection for PDAC within the Mayo Clinic Enterprise following NT between 2010–2022. No subjects received adjuvant therapy. All subjects achieved a minor pathologic response by CAP criteria defined as “partial response” or “no response” corresponding to CAP2 and CAP3, respectively. Variant exocrine carcinomas were excluded, in addition to those without research consent.

H&E Slide Digitization

A digital pathology library was created, consisting of H&E stained WSI of surgically resected PDAC. The most representative slides from each case (an average of five slides per case, range of 1–16) were selected. Each slide was scanned at 40X magnification (0.25 microns/pixel) via high-throughput digital imaging by Aperio GT 450.

PathExplore PDAC-Based Tissue and Cell Segmentation

To enable quantitative spatial analysis of the TME, we utilized PathExplore^™ PDAC (v1.0), a commercially available artificial intelligence (AI)-driven platform developed by PathAI, Inc. PathExplore PDAC is developed and evaluated using expert pathologist annotations on WSIs derived from pancreatic adenocarcinoma samples. PathExplore PDAC is designed for single-cell resolution spatial characterization of H&E-stained WSIs of PDAC specimens. It leverages deep learning models trained on pathologist-annotated datasets to identify and classify tissue compartments and cell types within the TME. PathExplore PDAC processed the digitized slides to perform automated tissue analysis, generating image overlays of cancer and cancer-associated stroma. Artifactual regions (e.g., folds, blur) were excluded via the PathExplore artifact model to ensure segmentation fidelity.

In this study, we specifically leveraged the spatial maps of cancer and stroma compartments to derive two-dimensional landscape mosaics. These served as the basis for TLA, a framework integrating landscape ecology metrics to quantify composition and configuration across the TME. PathExplore's harmonized tissue classes enabled consistent segmentation across samples, ensuring robustness for downstream statistical modeling. The segmentation results formed the foundation for extracting patch- and landscape-level metrics, which were subsequently aggregated at the patient level for association with disease-free survival (DFS).

Model performance for cancer and stroma segmentation within PathExplore PDAC was benchmarked using established precision and recall thresholds to assess agreement with pathologist annotations. For regions where the model-predicted area of cancer or cancer-associated stroma exceeded 5% of usable tissue, segmentation quality was classified as “Excellent” if both precision and recall exceeded 91%, “Good” if both exceeded 80%, and “Notable Over- or Underestimation” based on imbalanced precision-recall profiles or discrepancies with ground truth annotations. For predicted areas below 5%, a modified framework prioritized high recall and absence of false positives to mitigate misclassification of rare features. Only segmentation outputs meeting “Excellent” or “Good” thresholds were included for landscape metric extraction and subsequent analyses, ensuring high-confidence tissue compartment delineation across the cohort (Supplementary Fig. S1).

Model performance for lymphocyte classification was benchmarked against expert pathologist assessments, yielding an interclass correlation coefficient (ICC) of 0.84 in count of lymphocytes. This level of agreement is comparable to the inter-pathologist ICC observed among expert pathologists (Supplementary Table S1).

Tumor Landscape Analysis Pipeline

Spatial metrics were computed using Tumor Landscape Analysis (TLA), a pipeline of analysis adapted from landscape ecology methodologies^32,33 and recently described in a tumor landscape analysis framework (Cisneros et al., bioRxiv, 2025; DOI: 10.1101/2025.04.22.646608). For spatial modeling using TLA, input for the pipeline was derived from cancer and cancer-associated stroma overlays, two-dimensional landscape mosaics, provided by PathExplore PDAC. Each mosaic consists of a two-dimensional raster array of discrete categorical labels describing distinct regional patches, with each label corresponding to a tissue compartment class. Analysis was restricted to the cancer and stromal components, as these features have been linked to long-term survival outcomes.^10,11 Spatial metrics were computed at three hierarchical levels: the patch level, in which individual patches were quantified by their geometrical features, including area, edge length, and shape index; the landscape level, where whole-slide descriptors were generated based on the distributions of patch metrics and different measures of configuration and complexity of patch arrangements, including measures of diversity, entropy, and adjacency (interspersion); and the patient level, where each measurement across slides for each patient were summarized by either simple aggregation or the mean, minimum, and maximum of individual values across the slides when aggregation is not feasible.²⁰ A total of 42 spatial features for each patient were considered for our analysis. The Supplementary Methods document contains a glossary of terms and definitions of these features.

GATA6 IHC

GATA6 IHC was performed on 4 μm FFPE tumor sections using a GATA6 (D61E4) rabbit monoclonal antibody (Cell Signaling Technology, #5851, RRID:AB_10705521; 1:4000 dilution) on an automated staining platform, Leica BOND-RX (RRID:SCR_025548), with heat-induced epitope retrieval in EDTA buffer, pH 9.0. Staining was interpreted by two board-certified pathologists blinded to clinical outcomes. GATA6 expression was quantified using Histoscore (H-score) as previously described.³⁴ Briefly, the intensity of nuclear staining in positive tumor cells was assigned scores of 1+, 2+, and 3+, corresponding to weak, moderate, and strong staining, respectively. The proportion of positive tumor cells in each category was estimated, and a Histoscore was calculated (1 × % of cells with 1+ staining + 2 × % of cells with 2+ staining + 3 × % of cells with 3+ staining), yielding a total score ranging from 0 to 300 (Supplementary Fig. S2). Most cases had high GATA6 expression (>150 by H score) corresponding to the classical subtype. A cutoff of 100 was used keeping with the broad understanding that basal-type tumors have negative or weak GATA6 expression.

Outcomes Assessment

DFS was used as the primary endpoint of the study and is defined as the time from the date of curative resection to the date of first documented disease recurrence or death due to all causes, whichever occurred first.

Statistical Analysis

Spatial pathological features (SPFs) were summarized by mean (standard deviation [SD]), median, and range, compared by subgroups defined by clinical and disease characteristics. Two-level (high vs. low) risk classifications were developed based on the prognostic value of SPFs regarding DFS. Spearman correlation coefficients were used to describe the inter-correlations between SPFs to detect pairs with high collinearity. Normality transformations were performed when it is necessary to improve stability of the model fitting, by minimizing the influence of extreme values. Univariate association between individual SPF and DFS was assessed by Cox regression model, as well as spline fitting for non-linear effects. The individual SPFs with p-value < 0.20 were selected for subsequent modeling procedures for risk classification building. Two risk classification development methods were utilized. Recursive partitioning and tree regression method was used to split population into subgroups with homogeneity DFS which handles non-linear effects and interactions among SPFs naturally. High- vs. low-risk subgroups were defined by the end nodes generated by the tree method. The second method was based on traditional Cox regression with backward elimination procedure on selected SPFs after univariate analysis. The SPF variables with p value < 0.05 stayed in the final multivariable model. Interaction terms among remaining SPF variables were further evaluated. The final model was used to produce the continuous risk score for each patient. Since this score was associated with DFS in a linear form, median value was applied to define high-vs. low-risk groups. The discrimination abilities of both risk-classifications were quantified by C statistics. The risk-classifications were further evaluated by multivariable Cox models, adjusting for clinical and disease characteristics.

Code Availability Statement

All code for data processing and spatial analysis associated with this study is publicly available in the CarrLab GitHub repository (https://github.com/CarrLab/TLA). Documentation for usage, updates, and versioning is maintained at this location. The underlying TLA framework is described in detail in (Cisneros et al., bioRxiv, 2025; DOI: 10.1101/2025.04.22.646608).

Data Availability

The data generated in this study were collected under Mayo Clinic IRB 22–007646 and are subject to patient privacy protections that preclude unrestricted public release. De-identified data supporting the findings of this study, including patient-level spatial pathology features, clinicopathologic variables, and outcome data, are available from the corresponding author upon request, subject to a data use agreement and institutional approval. Whole slide images (WSIs) and the corresponding AI-derived tissue segmentation overlays contain potentially identifiable information and will be shared only under a formal data transfer agreement for non-commercial academic research use. Tissue and cell segmentation was performed using PathExplore PDAC, a commercial platform; classification thresholds and quality-control criteria used in this study are detailed in the Methods and Supplementary Data.

Results

Study Population, Clinicopathologic Features and Disease-Free Survival

The study cohort consists of 297 samples from 203 patients. Of these, 109 patients had only one WSI and 94 patients had two slides taken from adjacent cuts of the same tissue block. All slides were analyzed independently and then summarized across slides for each patient. In our cohort, 123 had PDAC recurrence, 60 of which had two WSIs included in the analysis. Of the 80 patients who remained recurrence-free during a follow-up period of up to 150 months, 34 had two WSIs analyzed. All 203 patients received NT. All patients in the present cohort achieved a minor pathologic response with College of American Pathologists (CAP) criteria of grades 2 or 3 indicative of partial response and no response, respectively. The mean follow-up time was 18.9 months. Demographic and clinicopathologic characteristics are summarized in Table 1 and Supplementary Table S2. The mean age at surgery was 65.4 years with a relatively balanced distribution of sex (47.3% female). Most patients had clinical stage IB (30.5%), IIB (25.5%), or III (26.0%) disease at diagnosis.

Table 1.

Demographics and Clinicopathologic Factors

	Overall (N=203)
Age at Surgery
Mean (SD)	65.4 (9.2)
Median (Range)	66.5 (39.4, 85.3)
Race
White	192 (95.5%)
Non-White	9 (4.5%)
Gender
Female	96 (47.3%)
Male	107 (52.7%)
Resectability 1
Resectable	90 (44.3%)
Borderline Resectable	60 (29.6%)
Locally Advanced	53 (26.1%)
Type of Surgical Resection
Distal Pancreatectomy	45 (22.2%)
Pancreaticoduodenectomy	126 (62.1%)
Total Pancreatectomy	32 (15.8%)
Histological Grade
1 - Well-differentiated	4 (2.0%)
2 - Moderately differentiated	138 (68.0%)
3 - Poorly differentiated	57 (28.1%)
4 - Undifferentiated	2 (1.0%)
X - Cannot be determined	2 (1.0%)
Pathologic Response (CAP)
2 - Partial Response	156 (82.2%)
3 - No Response	34 (17.9%)
Resection Outcome
R0	174 (86.1%)
R1	27 (13.4%)
R2	1 (0.5%)
Lymphovascular Invasion
Yes	27 (14.4%)
No	159 (84.6%)
Perineural Invasion
Yes	110 (57.3%)
No	80 (41.7%)
Pathologic T Stage
1a	7 (3.5%)
1b	4 (2.0%)
1c	35 (17.3%)
2	102 (50.5%)
3	44 (21.8%)
4	10 (5.0%)
Pathological N Stage
N0 (0 LN+)	140 (69.0%)
N1 (1–3 LN+)	45 (22.2%)
N2 (4+ LN+)	18 (8.9%)
Extent of Neoadjuvant Chemo
≥12 weeks	166 (85.6%)
<12 weeks	28 (14.4%)

^1.Per National Comprehensive Cancer Network criteria

Baseline resectability status (resectable, borderline resectable, locally advanced) was assigned using National Comprehensive Cancer Network criteria based on pretreatment cross-sectional imaging. 90 patients had resectable disease (44.3%), while 60 patients had borderline resectable disease (29.6%). Most subjects underwent pancreaticoduodenectomy (62.1%) and achieved an R0 resection (86.1%). The median time from diagnosis to surgery was 8.2 months (range 1.8–19.4 months). Histological grading indicated that 68.0% were moderately differentiated, while 28.1% were poorly differentiated.

Pathologic staging by AJCC v.8 criteria showed that 50.5% were classified as ypT2 tumors, while 21.8% and 5.0% were T3 and T4, respectively. Nodal involvement was observed at 31.0%, with 22.2% classified as N1 and 8.9% as N2. Perineural invasion (PNI) and lymphovascular invasion (LVI) were identified in 57.3% and 14.4% of cases, respectively. Most cases (80.9%) demonstrated high GATA6 expression (H-score ≥100), a validated immunohistochemical (IHC) surrogate^35,36 for the classical versus basal-like transcriptomic subtypes of PDAC.³⁷

Univariate analysis identified several clinicopathological factors associated with DFS (Supplementary Table S3). Pathologic T (ypT) stage (p<0.001) (Fig. 1A), pathologic N (ypN) stage (p=0.002) (Fig. 1B), and LVI (p=0.048) were significantly associated with DFS. Other factors, such as age, gender, histologic grade, and time from diagnosis to surgery, were not significantly associated with time to recurrence.

Figure 1. Clinicopathologic features associated with disease-free survival (DFS).

Kaplan-Meier curves stratifying DFS by (A) pathologic tumor (ypT) stage, (B) pathologic nodal (ypN) stage, (C) College of American Pathologists (CAP) tumor regression grade 2 (partial response) versus grade 3 (no response), (D) area of residual cancer within the tumor bed, and (E) ratio of cancer area to stroma area within the tumor bed.

While outcomes for cases with CAP grade 3 trended worse relative to those with CAP grade 2, this difference was not statistically significant (Fig. 1C). Given the subjective nature of CAP assessments by pathologists and risk of inter-observer variability^12,13, we assessed whether quantification of residual cancer could stratify outcomes in this cohort. Interestingly, quantifying the area occupied by cancer within the residual tumor was not associated with DFS (HR=1.12; 95% CI: 0.97–1.31; p=0.13) (Fig. 1D). The area occupied by cancer relative to the stroma area was also quantified much like published treatment response scores.¹³ This ratio was similarly not associated with outcomes (HR=1.06; 95% CI: 0.91–1.23; p=0.45) (Fig. 1E). While quantifying the amount of the residual tumor failed to stratify patient outcomes, this led us to explore whether the spatial composition and configuration of cancer and stromal components might offer more significant biological and clinical insights into residual disease. In addition, these findings underscore the heterogeneity of tumor characteristics and underline the persistent burden of disease in this cohort despite NT and surgical intervention.

Spatial Pathological Cohort Characterization

Using an artificial intelligence-enabled digital pathology platform, PDAC tissue overlays reduced to cancer and tumor-associated stroma patches, we summarized tissue composition (dominance and density of patches) and configuration (patch morphology and spatial arrangement) using 42 spatial features measured for each patient (see Supplementary Methods) and related these distributions to CAP treatment response categories (Fig. 2). Stromal patches were consistently larger on average and variable in size relative to cancer patches as measured by the mean and standard deviation (SD) of patch areas (Fig. 2A–B). When stratified by response, mean cancer patch area trended lower in CAP2 than CAP3, while variability in patch area separated groups for both stroma and cancer, with higher SD in CAP3 (Fig. 2A–B). Total region extent was captured by the Feret diameter (longest distance between any two points in a region)³⁸ of each class’s convex hull region (smallest convex polygon that overlaps all patches of the same class). The maximum and minimum values between slides for each patient were considered and, did not differ by response (Fig. 2C–D). Patch density differed by response for cancer but not for stroma, consistent with more numerous, smaller cancer islands in CAP3 (Fig. 2E). Edge density (the total patch interface length per unit area) was markedly higher in CAP3, indicating a more fragmented, interface-rich landscape when tumors failed to respond (Fig. 2F).

Figure 2. Cohort-level composition, configuration and diversity spatial features stratified by pathologic treatment response.

CAP response-stratified comparisons for spatial features. The plots show per-patient values summarized by CAP pathologic treatment response using violin plots with boxplot, rug marks, and overlaid points. CAP2 (partial response) is in red and CAP3 (no response) is in blue. Panels A-E represent tissue compartment-specific composition spatial features while panel F shows landscape edge density (a whole-slide composition metric). Panels G-J represent tissue compartment-specific configuration spatial features while panels K and L display whole-slide configuration metrics. Panel K is the mean adjacency index (homotypic contacts) and panel L is the mean contagion (landscape aggregation. Panels M and N show Shannon and Simpson diversity indices, respectively. SD = standard deviation; δ = Cliff’s delta.

Metrics of configuration also exhibited separation by treatment response. Fractal dimension, a boundary complexity measure that approaches a value of 1 for convex polygons and 2 for complex, plane-filling, dendritic shapes, showed higher mean stromal patch complexity and higher stromal variability in CAP3, though cancer fractal metrics were not discriminatory (Fig. 2G–H). However, there was some association between decreased fractal dimension and poorly differentiated disease (Supplementary Fig. S3) A similar metric, the mean shape index (a measure of shape complexity based on the relation between the area and the perimeter, see glossary section in Supplementary Methods) did not differ by response (Fig. 2I), but shape index variability, measured by the SD, did. Larger shape index variability, i.e. larger patch shape heterogeneity, was seen in CAP3 stroma and cancer (Fig. 2J), again suggesting tumor fragmentation as a signature of poor response to therapy. Adjacency index (frequency of like-with-like contact locations) and contagion (a measure of probability of adjacent locations belonging to the same class, see glossary section in Supplementary Methods) were both lower in CAP3, pointing to reduced homotypic aggregation and greater intermixing of cancer and stroma when treatment was ineffective (Fig. 2K–L). Finally, Shannon and Simpson diversity indices (see definitions in the Supplementary Methods) were elevated in CAP3, reinforcing the notion that non-responders harbor more compositionally and spatially diverse landscapes (Fig. 2M–N). Interestingly, neither of these two diversity measures are associated with DFS on their own (Table 2), and thus though they correlate with the initial response to therapy, they do not associate with long term survival of patients in this cohort. This observation suggests that spatial fragmentation might be a signature of initial tumor persistence after neoadjuvant therapy, but this characteristic by itself has no long-term effect in the progression of disease. None of these spatial features were consistently associated with common clinicopathologic features of disease such as differentiation or pathologic T or N staging (Supplementary Fig. S3–S5). Cohort-level distributions of composition/diversity and configuration spatial measures can be found in Supplementary Figures S6–S7.

Table 2.

Univariate Association Between Spatial Pathology Features and DFS (Cox Regression and Spline Models)

	Cox Model				Spline Analysis
Spatial Variables	HR	Conf.low	Conf.high	p-value	Overall p-value	Linear p-value	Non-linear p-value
ln(mean of cancer patch area)	0.93	0.62	1.38	0.7089	0.4375	0.8057	0.3813
ln(SD cancer patch area)	1.03	0.81	1.30	0.8069	0.7174	0.8396	0.6212
ln(mean stroma patch area)	1.27	1.03	1.56	0.0224	0.0385	0.0200	0.3638
ln(SD stroma patch area)	1.16	0.98	1.37	0.0929	0.1866	0.1231	0.5794
landscape patch density	0.99	0.87	1.13	0.8807	0.5252	0.9109	0.5822
tumor patch density	1.05	0.90	1.23	0.4986	0.3383	0.4407	0.3874
ln(stroma patch density)	0.79	0.64	0.97	0.0224	0.0385	0.0197	0.3681
ln(edge density)	1.40	1.04	1.87	0.0263	0.0636	0.0382	0.4208
mean cancer patch shape index	0.27	0.08	0.95	0.0407	0.1301	0.0383	0.6248
SD cancer patch shape index	0.81	0.25	2.62	0.7275	0.5199	0.7266	0.5609
ln(mean stroma patch shape index)	3.88	0.19	77.46	0.3754	0.1266	0.3265	0.1560
ln(SD stroma patch shape index)	2.38	1.26	4.49	0.0075	0.0383	0.0078	0.5265
Shannon diversity index	2.63	0.51	13.53	0.2477	0.5787	0.2729	0.7737
Simpson diversity index	2.39	0.24	23.58	0.4547	0.9353	0.4539	0.9810
min cancer Feret diameter	1.00	0.88	1.13	0.9733	0.7942	0.9910	0.7745
max cancer Feret diameter	0.99	0.86	1.14	0.8776	0.7579	0.8967	0.7358
ln(min cancer landscape shape index)	1.40	1.01	1.93	0.0433	0.1008	0.0682	0.4157
max landscape shape index	1.00	0.99	1.02	0.7825	0.7233	0.7862	0.7101

Univariate analysis is intended to reduce the number of features for subsequent modeling steps and not intended for final inference.

In sum, across scales, non-response was characterized by (1) higher variability in patch geometry (area, fractal dimension, shape index), most prominently within stroma, and (2) greater fragmentation and mixing (higher edge density, lower adjacency/contagion, higher spatial diversity).

Figure 3 presents a cohort-level heatmap of all spatial pathological features across individual patients. Two patient groups emerge when maximizing the average silhouette width for k-means³⁹ using all 42 features (also see Supplementary Fig.S8). Cluster 1 exhibits a fragmented, interface-rich phenotype characterized by a higher shape index, edge density and diversity indices, and lower contagion and homotypic adjacency. Cluster 2 shows the reciprocal pattern: a coarser, more aggregated tissue organization with less complex patch shapes. The corresponding clinical features are shown in color bars on top of the heatmap. Of these, CAP3 samples are significantly enriched in Cluster 1 (p-value < 0.0004), per a standard Chi-square goodness of fit test, while CAP2 samples are marginally enriched in Cluster 2 (p-value = 0.017). While Cluster 1 and Cluster 2 cases did not have statistically significant differences in DFS, these cohort-level patterns support the central premise that emergent spatial organization of the residual cancer–stroma ecosystem captures disease biology beyond gross burden alone. This motivated subsequent statistical modeling investigating the clinical significance of these spatial measures.

Figure 3. Unsupervised clustering of spatial pathology features identifies two residual tumor phenotypes.

A cohort-level heatmap of z-scored spatial pathology features (rows) across individual patients (columns). All 42 spatial features are presented, and 18 of them are intercorrelated (as described in Spatial Pathological Feature Selection section). Font colors, and numbers, in row labels indicate the correlated groups. All metrics with the same number cannot be discarded as equivalent when considering Pearson’s correlation confidence interval. Rows are sorted by the difference between the median value of the metric across patients in Cluster 1 versus Cluster 2. The color scale reflects relative deviation from the cohort mean. CAP = College of American Pathologists pathologic treatment response; ypN = pathologic lymph node status; ypT = pathologic tumor status; DFS = disease free survival.

Spatial Pathological Feature Selection

Cross-correlation between all 42 variables reveals a significant overlap for 24 of them, which group into 6 different clusters. Each cluster is comprised of 2 to 7 variables for which the confidence interval of Pearson’s cross-correlation overlaps the value 1.0, indicating that full correlation cannot be statistically rejected, and thus they are likely to carry the same information (correlogram is shown in Supplementary Fig. S9). For each of the clusters, the variable that has the lowest p-value in a Wilcox test that distinguishes the recurrence status was chosen as the best representative. In addition, a total of 18 variables did not cluster with any others, thus the final set considered for statistical modeling contains 24 variables. These variables were then evaluated to identify those potentially associated with DFS. Screening univariate survival analyses identified seven features with p-values <0.20: natural log (ln) of the mean stroma patch area (the mean area of all stroma patches), ln(SD stroma patch area), ln(stroma patch density), ln(edge density), mean cancer patch shape index (the mean shape index of all cancer patches, CSI), ln(SD stroma patch shape index (the SD of shape indices of all stroma patches, SSI)), and the ln(minimum cancer landscape shape index) (minimum landscape index between slides for each patient) (Table 2). To further evaluate the potential biological and clinical significance of these features, statistical risk models were developed.

Variability in Stromal Shape Index and Mean Cancer Shape Index (Recursive Partitioning and Tree Regression Model)

Using recursive partitioning and tree regression (RPTR) to stratify patients into high- and low-risk groups, CSI and SSI were identified as significantly discriminatory (Supplementary Table S4). Shape index is a metric used to measure the complexity of the shape of a habitat patch, providing insight into how irregular or convoluted the boundary of the patch is compared to a standard geometric shape, such as a square or circle. Shape index is calculated as follows:

$$SI=\frac{p}{4\sqrt{a}}$$

Where p represents the perimeter of the patch, a is the area of the patch. A shape index of 1 is indicative of a perfectly square patch, representing minimum complexity. Alternatively, the greater the shape index, the more irregular or elongated the patch.

Risk classification was specifically defined by ln(mean CSI) (Fig. 4A and B) and ln(SD SSI) (Fig. 4C and D). High-risk disease was defined as ln(mean CSI) ≤ 1.775 and ln(SD SSI) > 0.651. High-risk patients (n=63) had a median DFS of 7.23 months (95% CI: 5.19–9.63) compared to 11.57 months (95% CI: 9.24–17.92) for low-risk patients (HR=1.82; 95% CI: 1.31–2.52; p=0.0003) (Fig. 4E). After adjusting for the ypT and ypN stages, PNI/LVI, resection outcome, local invasion, and chemotherapy given by multivariable Cox regression analysis (Table 3), spatial indices remained independently prognostic for DFS. High-risk classification was associated with a 1.71-fold increased risk of recurrence (HR=1.71, 95% CI: 1.17–2.63, p=0.003). Sensitivity analysis also included other clinicopathologic covariates, which resulted in consistent results (Supplementary Table S5).

Figure 4. RPTR Model: Shape index of residual cancer and stroma in post-treatment PDAC and associated DFS.

(A, B) Representative cases illustrating high and low mean cancer shape index (CSI), respectively. (C, D) Representative cases illustrating high and low standard deviation of stroma shape index (SSI), respectively. Each panel (A–D) displays, from left to right: hematoxylin and eosin (H&E)–stained whole-slide images at 1× magnification, corresponding AI-generated overlays with cancer (purple) and stroma (yellow) segmentation, views at 10× magnification (for A, B), and a graphical representation of the spatial metric. Shape index (SI) is calculated as SI = P/(2√πA), where P is the patch perimeter and A is the patch area. A value of 1 indicates a perfectly circular shape, while higher values reflect increasing boundary irregularity. (E) Kaplan–Meier plot showing DFS stratified by risk group, defined by ln(mean CSI) and ln(SD SSI).

Table 3.

Multivariable Cox Regression Analyses of Disease-Free Survival for the RPTR and Cox Spatial Risk Models.

Characteristic	RPTR Model (Cancer Mean CSI + Stromal SD SSI)			Cox Model (Mean Stromal Patch Area + Edge Density)
	HR	95% CI	p-value	HR	95% CI	p-value
Risk Category
Low Risk (reference)	—	—		—	—
Middle Risk				1.43	0.93, 2.20	0.11
High Risk	1.71	1.21, 2.43	0.003	2.19	1.34, 3.58	0.002
Pathologic T-Stage
T1 (reference)	—	—		—	—
T2	1.16	0.74, 1.81	0.5	1.20	0.77, 1.86	0.4
T3	2.24	1.35, 3.71	0.002	2.06	1.23, 3.42	0.006
T4	2.30	1.00, 5.26	0.049	2.48	1.08, 5.71	0.032
Pathologic Nodal Involvement
Negative (reference)	—	—		—	—
Positive	1.76	1.25, 2.49	0.001	1.70	1.20, 2.40	0.003
Neoadjuvant Chemotherapy Regimen
First-Line GnP (reference)	—	—		—	—
First-Line FFX	0.87	0.56, 1.35	0.5	0.93	0.60, 1.45	0.8
First-Line FFX, Second-Line GnP	1.12	0.69, 1.81	0.7	1.18	0.73, 1.92	0.5
Model Statistics	c-index = 0.632; Log-likelihood = −658 No. obs. = 193; N events = 148			c-index = 0.644; Log-likelihood = −658 No. obs. = 193; N events = 148

Abbreviations: HR = Hazard Ratio, CI = Confidence Interval; CSI = Cancer Shape Index, SSI = Stroma Shape Index, GnP = gemcitabine + nab-paclitaxel, FFX = FOLFIRINOX.

Mean Stromal Area and Edge Density (Cox Proportional Hazards Model)

An alternative statistical model employing backwards elimination (Supplementary Table S6) identified edge density (Fig. 5A and B) and mean stroma area (Fig. 5C and D) as significant spatial features. While the mean stroma area is self-explanatory, edge density quantifies the total length of all edge segments in a landscape, standardized by the total area of the landscape or patch, providing an understanding of landscape fragmentation and complexity. Edge density is calculated as follows:

$$ED=\frac{E}{A}$$

Figure 5. Cox Model: Edge density and mean stroma area in post-treatment PDAC and associated DFS.

(A, B) Representative cases illustrating high and low edge density, respectively. (C, D) Representative cases illustrating high and low mean stroma area, respectively. Each panel (A–D) displays, from left to right: H&E–stained whole-slide images at 1× magnification, corresponding AI-generated overlays with cancer (purple) and stroma (yellow) segmentation, views at 10× magnification (for A, B), and a graphical representation of the spatial metric. (E) Kaplan–Meier plot showing DFS stratified by quartiles on a prognostic index derived from mean stroma area and edge density. (F) Prognostic refinement using an integrated model incorporating ln-transformed cancer area, mean stroma area, and their interaction, demonstrating improved risk discrimination.

Where E represents the total length of edges and A is the total area of the landscape. Higher edge density values indicate greater fragmentation, while lower values are suggestive of a more compact, continuous habitat with fewer boundaries.

The prognostic index was calculated as 0.2701 * ln(mean stromal patch area) + 0.3905 * ln(edge density). High-risk patients (n=51) had a median DFS of 6.21 months (95% CI: 4.64 – 9.86), intermediate-risk patients (n=101) had a median DFS of 9.37 months (95% CI: 8.05 – 15.62), while low-risk patients (n=51) had a median DFS of 20.91 months (95% CI: 10.85 – 33.93) (Fig 5E). Multivariate Cox regression analysis confirmed that these spatial features are independent risk factors for DFS, with high-risk classification associated with a 2.19-fold higher risk of recurrence (HR=2.19, 95% CI: 1.21–3.10, p=0.002) (Table 3). Sensitivity analysis also included other clinicopathologic covariates, which resulted in consistent results (Supplementary Table S7).

Integration of Cancer Area

Extensive literature underscores that the extent of residual cancer, often subjectively assessed by tumor area or burden, holds independent clinical significance. To align our risk model with these historical insights, we performed an exploratory analysis incorporating the cancer area as an additional feature in the modeling process. In univariate analysis, the natural log-transformed cancer area demonstrated a p-value of 0.178 for association with DFS, meeting our threshold (p<0.2) for variable advancement.

To evaluate the impact of integrating cancer area alongside spatial features, we replaced edge density with ln(cancer area) in RPTR models, as well as in Cox proportional hazards models. Although RPTR models remained unchanged from prior analyses, backward selection approaches, which forced ln(cancer area) into the model, identified a prognostic combination of ln(cancer area), ln(mean stroma patch area), and ln(edge density).

Further refinement through interaction modeling revealed that the combination of ln(cancer area) and ln(mean stroma patch area), including their interaction term, significantly stratified patients into distinct risk categories for DFS (Fig. 5F). This integrated model, adjusting for covariates, demonstrated improved discriminatory ability, with C-statistics reaching 0.666, supporting the complementary, rather than independent prognostic value of cancer area alongside established spatial features. Collectively, these analyses validate the inclusion of cancer area in computational pathology-derived risk models, providing a framework that incorporates both tissue architecture and tumor burden as clinically important features in residual PDAC.

Spatial Risk Models Stratify Disease-Free Survival within Pathologic Stage

Because pathologic T and N stage are associated with DFS, and both spatial models remained prognostic after adjustment for stage, we evaluated whether spatial risk further stratified outcomes within pathologic stage groups. In the RPTR model, joint stratification by spatial risk and pathologic T stage separated patients into four DFS trajectories (Fig. 6A, log-rank p<0.001). Patients with low-risk pT1/2 disease had the most favorable outcomes, whereas those with high-risk pT3/4 disease had the poorest outcomes. RPTR risk also separated DFS across nodal strata (Fig. 6B, log-rank p<0.001). The Cox spatial risk model showed a similar pattern when high- and low-risk patients were stratified by pathologic T stage (Fig. 6C, log-rank p=0.001) and pathologic N stage (Fig. 6D, log-rank p<0.001). Together, these stratified analyses indicate that residual tumor-stroma spatial organization provides prognostic information complementing pathologic staging.

Figure 6. Spatial risk classification stratifies disease-free survival across pathologic T and N stages.

Kaplan-Meier curves of disease-free survival in patients jointly stratified by spatial risk classification and pathologic stage. (A) RPTR risk (High vs. Low) and pathologic T stage (pT12 vs. pT3/4); n = 202. (B) RPTR risk and pathologic N stage (pN0 vs. pN1/2); n = 203. (C) Cox model risk and pathologic T stage; n = 101. For clarity, only patients classified as either high or low Cox risk are presented. (D) Cox model risk and pathologic N stage; n = 102. Same subset as in (C). Log-rank p values comparing the four strata within each panel are displayed.

Complementarity of Spatial Risk Models

The RPTR and Cox models share no features and were derived using different statistical methods. Because the upstream feature-reduction step removed cross-correlated redundancy across all candidate spatial variables (Supplementary Fig. S9), the non-overlapping variables selected by the two models reflect distinct architectural information rather than alternative measurements of the same quantity. We therefore examined whether the two classifications stratify the same patients. Cross-classification (Supplementary Fig. S10, Supplementary Table S8) showed that the two models broadly converged on patients at the extremes of risk while disagreeing meaningfully in between: 52% of RPTR high-risk patients were also classified as Cox high-risk and 6.3% as Cox low-risk, while 34% of RPTR low-risk patients were classified as Cox low-risk and 13% as Cox high-risk. Base-model discrimination was comparable (C-statistics of 0.632 for RPTR and 0.644 for Cox). Thus, the models each interrogate a distinct architectural dimension of the post-treatment tumor.

Lymphocyte Spatial Patterns Associated with Risk Model Pathologic Features

While the primary aim of the present study is to evaluate the clinical importance of tumor topology, a natural extension of this work is to explore the cellular correlates that may underlie these spatial features. To this end, we examined the relative spatial distribution of tumor-infiltrating lymphocytes (TILs) across the tissue compartments defined by our segmentation framework and compared these distributions between the risk groups derived from each survival model.

Three regional zones were considered: intratumoral (cancer patches), peritumoral (a 20 μm annular zone surrounding cancer patch boundaries), and stroma (stromal patches). For each zone, two TIL metrics were quantified: lymphocyte density (cells/mm²) and lymphocyte infiltration ratio (the proportion of lymphocytes residing in each zone relative to the total in the whole slide). These metrics were compared across risk groups for both the RPTR model (CSI and SD SSI) and the Cox model (edge density and mean stromal patch area).

For the RPTR model, high-risk patients had significantly lower intratumoral lymphocyte density and a lower intratumoral infiltration ratio compared to low-risk patients, indicating that high-risk tumors tend toward an immune-excluded or immune-cold intratumoral phenotype (Fig. 7A, bottom and top rows, intratumoral). In the peritumoral zone, the infiltration ratio was significantly higher in high-risk patients, while lymphocyte density did not differ significantly between groups, suggesting that lymphocytes are proportionally concentrated at the tumor periphery in high-risk cases without a net increase in absolute immune cell number at that interface (Fig. 7A, peritumoral). Stromal lymphocyte density and infiltration ratio were both significantly higher in high-risk patients, consistent with lymphocyte sequestration within the stroma rather than intratumoral penetration (Fig. 7A, stroma).

Figure 7. Tumor-infiltrating lymphocyte (TIL) distributions across spatial risk groups and their correlations with model-defining spatial features.

(A) TIL metrics across risk groups defined by the mean cancer shape index (CSI) and SD stroma shape index (SSI) RPTR model, stratified by tissue zone (intratumoral, peritumoral, stroma). The top row shows lymphocyte infiltration ratio (proportion of total lymphocytes per zone); the bottom row shows lymphocyte density (cells/mm²). Each violin plot displays the full distribution with an embedded boxplot (median and IQR) and mean (triangle). Statistically significant pairwise comparisons are annotated with p-values and Cliff's δ effect sizes. (B) Scatter plots showing Pearson and Spearman correlations between TIL metrics (infiltration ratio, top; lymphocyte density, bottom) within the intratumoral zone and the RPTR model features: mean CSI (left) and SD SSI (right). (C) TIL metrics across risk groups defined by the edge density and mean stromal patch area Cox proportional hazards model, displayed using the same format as panel A. (D) Scatter plots of Pearson and Spearman correlations between TIL metrics and the Cox model features: edge density (left) and mean stromal patch area (right). SD = standard deviation; IQR = interquartile range.

Correlation analysis confirmed monotonic relationships between the RPTR spatial features and TIL metrics (Fig. 7B). Higher SD SSI was negatively correlated with both intratumoral lymphocyte density and infiltration ratio, while higher mean CSI showed a modest positive association with lymphocyte density but not with infiltration ratio.

For the Cox model, risk group differences in TIL metrics were more pronounced across all three zones and all pairwise comparisons, with effect sizes ranging from moderate to large (Fig. 7C). High-risk patients again demonstrated the lowest intratumoral lymphocyte density and ratio, with stepwise gradation across intermediate and low-risk groups. Peritumoral differences were particularly striking, with the largest effect sizes observed across pairwise comparisons, reinforcing the pattern of perilesional immune accumulation in higher-risk disease. Correlation analyses with Cox model features showed that larger mean stromal patch area was consistently and negatively associated with both TIL density and infiltration ratio, while edge density showed weaker and directionally mixed associations (Fig. 7D).

Collectively, these findings indicate that the spatial risk configurations identified by our models correspond to distinct immune infiltration patterns: high-risk landscapes are characterized by intratumoral immune exclusion with relative lymphocyte accumulation at the tumor periphery and in the stroma, a distribution consistent with active immune sequestration. Whether this immune redistribution is a cause or a consequence of the fragmented, interface-rich topologies that define high-risk disease remains to be determined, but these observations provide a cellular correlate for the spatial risk signal and motivate future mechanistic studies with higher-resolution immune phenotyping.

Discussion

This study demonstrates that in PDAC treated with NT and resection, but limited by minor pathologic response, the spatial architecture of the residual tumor-stroma ecosystem carries independent, clinically meaningful information about recurrence risk that is not captured by conventional response grading. Specifically, two multivariable spatial models including (1) a configuration model defined by mean CSI and variability of SSI and (2) a composition model defined by mean stromal patch area and landscape edge density, were independently associated with DFS after adjustment for established clinicopathologic covariates. Notably, neither CAP2 vs CAP3 response scoring, nor the area of residual cancer alone stratified DFS in this minor-response cohort, underscoring a gap in current assessment paradigms that our spatial approach begins to fill.

These results both refine and extend prior observations that major pathologic response is favorable in PDAC, whereas patients with residual disease show heterogeneous outcomes despite similar CAP grades^10–15. Traditional grading emphasizes “how much” tumor remains and is vulnerable to inter-observer variability^12–15. Our analysis asks a different question: how is the residual disease arranged? By quantifying patch shapes, sizes, densities, and the amount of cancer-stroma interface, we show that organization, not just extent, encodes treatment-resistant biology after NT.

This work has limits. It is a retrospective, single-enterprise study (Supplementary Table S9). A key limitation is the absence of paired pre-treatment specimens, which precludes direct within-patient assessment of therapy-induced remodeling of spatial configurations. In our practice, most pre-treatment diagnostic specimens are fine needle aspirations that do not preserve sufficient architectural integrity for reliable quantification of tumor-stroma topology. Future studies incorporating paired core biopsies or other serial sampling strategies, when feasible, will be important to directly test spatial remodeling over time. External and prospective validation is needed across scanners and staining protocols. We also assume that the one-to-two WSIs evaluated per case are representative of the resected tissue. Although quality controls were applied, any AI-assisted pipeline can inherit biases from training data and class definitions. The TIL analyses presented here represent an initial step toward resolving the cellular constituents that underlie spatial risk signatures; however, characterization of fibroblast subtypes, vascular populations, and functional immune states at and across edge habitats remains an important direction for future work. Prospective integration and validation of spatial metrics into adjuvant decision workflows are important next steps.

Nonetheless, viewed through an ecological lens (Cisneros et al., bioRxiv, 2025; DOI: 10.1101/2025.04.22.646608), several clear patterns emerged in our findings. First, greater fragmentation and mixing (higher edge density, higher diversity, and lower homotypic aggregation) were hallmarks of non-response at the cohort level and signaled worse DFS in multivariable models. Second, shape features mattered: lower mean CSI (more compact cancer patches) together with higher variability in SSI (more heterogeneous stromal shapes) defined a high-risk configuration class. These observations align with mechanistic ideas that cancer–stroma interfaces act as privileged habitats for nutrient exchange^18–20, immune evasion or exclusion ^21–25, and mechanically favorable routes for invasion and collective migration^26–30. The architecture of the resection specimen may reflect microenvironmental contexts associated with persistence after NT, though causality cannot be inferred without serial sampling. Interface-rich, intermixed landscapes create more edge niches and a wider array of microenvironments for selection to exploit, so that any microscopic or occult residual disease is more likely to reside in resource-rich, immune-protected, mechanically permissive habitats. Although one might predict that increased interface density would enhance immune surveillance, prior work in PDAC suggests that stromal and matrix populations at these borders can be predominantly immunosuppressive and barrier-forming, shielding rather than exposing cancer cells^40–44. The TIL analyses in Figure 7 provide direct evidence consistent with this interpretation. High-risk spatial configurations were marked by reduced intratumoral lymphocyte density and infiltration ratio, with relative TIL accumulation at the periphery of cancer patches and within the stroma, a distribution consistent with immune sequestration rather than intratumoral infiltration. While our landscape framework treats stroma as a single compartment and therefore cannot resolve fibroblast, vascular, or functional immune subsets, the coupling of spatial topology with lymphocyte distribution implies that the organization features we quantify carry biologically consequential immune correlates.

The clinical message is pragmatic. First, post-NT risk stratification improves when quantitative spatial metrics are added to standard pathology. Our CSI/SSI and stroma-area/edge-density models outperformed CAP grading in this setting and remained independent of ypT, ypN, margin status, and PNI/LVI. These signatures could help identify patients for intensified or mechanism-matched adjuvant strategies even when traditional indicators are equivocal. Second, the metrics are observer-independent and automatable on routine H&E slides, making them suitable for real-world deployment and for risk adaptive trials that already pivot on residual disease but currently lack spatial resolution^45–47. The immune exclusion pattern observed in high-risk patients, characterized by intratumoral TIL depletion and relative peritumoral and stromal sequestration, provides a cellular rationale for prioritizing immune-access strategies in this subgroup. Specifically, patients with high edge density and high stromal-shape variability, who also demonstrate this immune-excluded phenotype, may represent candidates for approaches designed to overcome stromal immune barriers, such as stroma-modulating agents, matrix and biomechanics pathway inhibitors, or immunotherapies that target the interface microenvironment. By contrast, patients with coarser, more aggregated landscapes may warrant distinct mechanistic strategies. Treatment-specific interactions and prospective integration into adjuvant decision workflows remain important next steps.

Taken together, reframing treated PDAC as a post-therapy landscape reveals that the topology of residual cancer-stroma, not merely its quantity, tracks with therapeutic resistance and risk of relapse. Interface-rich fragmentation, compact cancer geometry, and heterogeneous stromal shapes define high-risk ecologies independent of standard features, whereas coarser, less interspersed architectures associate with longer DFS and may reflect a more limited repertoire of surviving niches. Critically, these spatial risk signatures are coupled to a distinct immune infiltration phenotype with high-risk topologies corresponding to intratumoral immune exclusion with relative lymphocyte sequestration, providing a cellular dimension to what has otherwise been a purely architectural risk signal. The spatial patterning we observe in the resected specimen may be consistent with microenvironment contexts linked to persistence and recurrence risk, motivating future studies that directly interrogate cellular states at interfaces and, when feasible, incorporate pre-/post-treatment sampling. These insights motivate spatially informed, risk-adaptive adjuvant strategies and mechanistic studies that resolve the cellular constituents and signaling pathways at edge habitats and mixing zones as therapeutic vulnerabilities, with the broader aim of converting descriptive histology into actionable ecology for patients who need improved post-NT outcomes.

Translational Relevance.

Pancreatic ductal adenocarcinoma (PDAC) frequently recurs after neoadjuvant therapy and curative-intent resection, with most patients showing only minor pathologic response and variable outcomes. In 203 resected PDAC cases with minor response, we applied an AI-enabled digital pathology workflow to routine H&E whole-slide images segmenting cancer and stroma and to computing quantitative spatial metrics predictive of disease-free survival. A fragmented, interface-rich, intermixed tumor-stroma landscape with compact cancer geometry and heterogeneous stromal shapes identified patients with significantly shorter disease-free survival, independent of standard clinicopathologic factors. By contrast, cancer area and pathologist treatment response assessment did not reliably stratify risk. Two complementary multivariable spatial risk models, applied to final surgical pathology, improved predictions and can be automated on routinely generated slides without additional assays. These findings support spatial pathology of resected post-neoadjuvant specimens for adjuvant risk stratification and motivate prospective validation in independent cohorts, and treatment-naïve upfront resected specimens.

Abbreviations:

AI

artificial intelligence

AJCC

American Joint Committee on Cancer

CAP

College of American Pathologists

CI

confidence interval

CSI

cancer shape index

DFS

disease-free survival

EDTA

ethylenediaminetetraacetic acid

FFX

FOLFIRINOX

FFPE

formalin-fixed paraffin-embedded

GnP

gemcitabine plus nab-paclitaxel

H&E

hematoxylin and eosin

HR

hazard ratio

ICC

intraclass correlation coefficient

IHC

immunohistochemistry

IQR

interquartile range

IRB

Institutional Review Board

LN

lymph node

LVI

lymphovascular invasion

NT

neoadjuvant therapy

PDAC

pancreatic ductal adenocarcinoma

PNI

perineural invasion

RPTR

recursive partitioning and tree regression

RT

radiation therapy

SD

standard deviation

SI

shape index

SPF

spatial pathological feature

SSI

stroma shape index

TIL

tumor-infiltrating lymphocyte

TLA

Tumor Landscape Analysis

TME

tumor microenvironment

WSI

whole-slide image

ypN

post-treatment pathologic nodal stage

ypT

post-treatment pathologic tumor stage