Source-linked AI summary
A Multimodal Dataset for Survival Prediction in Resected Pancreatic Ductal Adenocarcinoma
Anh-Tien Nguyen, Mawuko Tettey, Jacqueline Michelle Metsch, Teresa Zimmer, Niklas Ullrich, Mario Duker, Sandra Rungeling, Kirsten Reuter-Jessen, Tessa Rosenthal, Lena-Christin Conradi, Michael Ghadimi, Alexander Konig, Elisabeth Hessmann, Volker Ellenrieder, Philipp Strobel, Hanibal Bohnenberger, Anne-Christin Hauschild
TL;DR
Survival research in resected PDAC is limited by the scarcity of datasets linking histology with clinical, molecular, and long-term outcome data. This study presents a linked institutional cohort and evaluates multimodal and unimodal survival-prediction configurations. Ridge Cox regression with numeric clinicopathological and mutation variables achieved the highest mean concordance overall, while multimodal fusion led among neural models; these are initial benchmarks for future work.
Problem
Pancreas-specific survival research has few datasets linking whole-slide histology with clinical follow-up and molecular data at a survival-relevant scale.
Source: §1 · pp. 1–2Method
The study assembled a resected PDAC cohort linking multiple H&E WSIs with clinical variables, targeted mutation data, and overall-survival outcomes, then benchmarked fourteen prediction configurations.
Source: §6 · p. 11Results
C-index was 0.652 ± 0.046 for ridge-regularised Cox regression with numeric clinicopathological and mutation variables, 0.603 ± 0.030 for image-only attention, and 0.619 ± 0.025 for multimodal co-attention.
Source: §6 · p. 11Takeaways & Limitations
The dataset and benchmarks provide a pancreas-specific foundation for future survival research and cross-cohort validation, as initial reference values rather than a definitive modeling strategy.
Source: §6 · p. 11
Abstract
from arXiv · showhide
Survival research in pancreatic ductal adenocarcinoma (PDAC) is limited by the scarcity of datasets linking whole-slide histology with clinical, molecular, and long-term outcome data. We present a retrospective single-centre cohort of 302 patients who underwent PDAC resection at University Medical Center Gottingen. The dataset comprises 446 H&E whole-slide images, clinicopathological variables, targeted sequencing data for 154 patients, and overall-survival outcomes. During follow-up, 253 patients died, and the median follow-up was 76 months. To establish initial reference values, we evaluated fourteen survival-prediction configurations using identical five-repetition Monte Carlo cross-validation partitions. Ridge Cox regression using numeric clinicopathological variables achieved a mean concordance of $0.649 \pm 0.042$ and $0.652 \pm 0.046$ after adding KRAS and TP53 mutation status. The image-only attention model achieved $0.603 \pm 0.030$, while multimodal fusion achieved $0.619 \pm 0.025$, the highest concordance among the neural models. These results establish promising initial benchmarks for future research using this pancreas-specific multimodal dataset, paving the way for external validation.
1 Introduction
The paper motivates individualized prognosis in resected pancreatic cancer because stage-based assessment misses biological heterogeneity and pancreas-specific multimodal survival data are scarce. It introduces a linked histology, clinical, and molecular dataset with initial overall-survival benchmarks.
1 Introduction
- Reliable individualized prognosis is important for treatment decisions, clinical-trial identification, and post-surgical follow-up, while pancreatic cancer survival remains below 15%.The introduction also links the prognosis gap to late-stage diagnosis and difficulty identifying resected patients likely to benefit from adjuvant therapy or at high risk of early recurrence.Source: §1 · pp. 1–2
- Stage-based prognosis has wide survival overlap and misses biologically distinct tumor subtypes, motivating joint modeling of histology, clinical covariates, and molecular measurements.The introduction notes that resected pancreatic cancer prognosis still primarily relies on TNM stage and a small set of pathology variables.Source: §1 · p. 2
- Pancreas-specific multimodal survival modeling is constrained by scarce linked histology, clinical follow-up, and molecular data, alongside low tumor cellularity and intratumoral morphological heterogeneity.Prior pancreas-specific work has largely analyzed single modalities or included small pancreatic subsets in pan-cancer benchmarks.Source: §1 · p. 2
- The study introduces a multi-slide pancreatic cancer cohort with matched clinical variables and targeted mutation sequencing, then benchmarks shared survival models against unimodal and pairwise-modality baselines.The authors frame the results as initial reference values and the dataset as a reproducible pancreas-specific test set for future research.Source: §1 · p. 2
2 Related Work
Public pancreatic histopathology resources are fragmented, and the principal cohorts combining WSIs, molecular profiles, and survival each include fewer than 200 patients. Prior computational pathology work spans diagnosis, molecular characterization, prognosis, and tissue quantification, while this study contributes an independent multimodal PDAC cohort and initial benchmarks.
2.1 Public histopathology resources for pancreatic cancer
- TCGA-PAAD and CPTAC-PDA are the only public cohorts identified as combining pancreatic cancer WSIs, molecular profiles, and survival, and neither provides a large independent population for development and external validation.TCGA-PAAD and CPTAC-PDA each contain fewer than 200 patients; other resources are fragmented across tasks or lack longitudinal outcomes.Source: §2.1 · p. 3
2.2 Computational pathology of pancreatic cancer
- Pancreatic computational pathology spans diagnosis, molecular characterization, prognosis, and tissue quantification, but larger development and validation cohorts generally remain institutional or consortium resources that are not publicly redistributable.Diagnosis models can show substantial cross-cohort performance differences, including an internal AUC of 0.997 versus 0.668 on CPTAC in one example.Source: §2.2 · pp. 3–4
- Pan-cancer foundation-model evaluations include pancreatic data, but dedicated pancreas-specific survival validation remains limited.UNI, CONCH, and Prov-GigaPath did not report dedicated pancreatic benchmarks in their original evaluations; PRISM2 included PAAD as one of 13 disease-specific survival tasks.Source: §2.2 · p. 4
- The study introduces an independent institutional PDAC dataset linking multiple H&E WSIs per patient with clinical, targeted sequencing, and long-term survival data, and reports initial modality-combination benchmarks.The benchmarks are intended as cohort reference performance rather than a definitive survival model, supporting future cross-cohort validation and multimodal research.Source: §2.2 · p. 4
3 Dataset
The dataset is a retrospective, single-centre cohort of resected PDAC patients with digitized histology, clinical covariates, targeted genomic profiling, and long-term survival follow-up. Outcomes are represented on a discrete time grid, and models use repeated patient-level cross-validation with shared partitions.
3.1 Study population and data sources
- Of 631 screened patients, 302 met eligibility criteria for the resected PDAC cohort; 253 died during follow-up, and median follow-up was 76.0 months.Patients with neoadjuvant treatment were excluded, and eligible patients required documented overall survival and at least one archival H&E WSI.Source: §3.1 · pp. 4–5
3.2 Clinical variables
- The primary clinical feature set comprised seven routine variables, while targeted sequencing covered 102 cancer-associated genes and KRAS and TP53 status was defined from pathogenic or likely pathogenic variants.Missing clinical covariates were uncommon and imputed using medians from the corresponding training split.Source: §3.2 · pp. 5–6
3.3 Outcome encoding for individual survival distributions
- Overall-survival outcomes were encoded as a time-bin index and death or censoring indicator, with eleven intervals spanning six-month periods through 60 months and a final open-ended interval.For censored patients, the label indicates survival up to censoring rather than death within the assigned interval.Source: §3.3 · pp. 6–7
3.4 Cross-validation design
- Five Monte Carlo cross-validation repetitions used patient-level training, validation, and test partitions, and the same stored partitions were reused across all models.Each repetition split 302 patients into 152 training, 75 validation, and 75 test cases; test sets could overlap between repetitions.Source: §3.4 · p. 7
4 Benchmark
The benchmark compares input configurations, encoders, and survival models across clinicopathological variables, mutation status, and WSI data. It uses concept-guided patch selection, Cox and neural survival approaches, and common held-out evaluation across repeated splits.
4.1 Overview
- Fourteen configurations compared numeric covariates, text encodings of the same covariates, WSI features, and multimodal image-text fusion using shared patient splits and common outcome and metric definitions.The design separated input configuration, encoding backbone, and survival model; ridge Cox and a deep survival network on numeric covariates served as baselines.Source: §4.1 · p. 7
4.2 Input representations
- Seven routine clinicopathological variables, with optional KRAS and TP53 indicators, were encoded numerically or as templated clinical text; frozen BioClinical ModernBERT and CONCH text encoders produced text representations.WSI patches were encoded with the frozen CONCH image encoder, whose image and text encoders share an embedding space.Source: §4.2 · pp. 7–8
4.3 Whole-slide token selection
- The deterministic FOCUS procedure selected and pruned WSI patches using contrast scores against 22 pancreatic pathology text prompts and a neighboring-patch embedding-similarity threshold of 0.9.The prompt bank included 14 tumor-related and 8 background concepts.Source: §4.3 · p. 8
4.4 Survival models
- The attention-based WSI model used tumor-concept queries to attend over selected patch tokens, then pooled the concept-specific representations for discrete-time survival prediction.The discrete-time head mapped encoded features to eleven logits, transformed into interval hazards and an individual survival distribution.Source: §4.4 · pp. 8–9
- Multimodal co-attention used nine CONCH-derived clinical and mutation text tokens as queries over WSI tokens, then separately pooled image-conditioned and text representations for survival prediction.The implementation adopted the co-attention component of MCAT but not its original transformer encoders.Source: §4.4 · p. 9
4.5 Evaluation and statistical analysis
- Harrell’s concordance index was calculated on each held-out test set and summarized as mean ± standard deviation across five repetitions; models also generated patient-specific survival probabilities at six-month evaluation points.Concordance evaluates risk-score ranking among patient pairs comparable under right censoring.Source: §4.5 · p. 9
5 Results
Across fourteen configurations, numeric clinicopathological variables with TP53 and KRAS status gave the highest mean concordance overall, while multimodal co-attention led the neural models. Results are initial cohort benchmarks, not a definitive modeling strategy.
5 Results
- 0.649±0.042 for ridge-regularised Cox regression and 0.599±0.040 for the deep discrete-time model were clinical-feature references, but their fitting protocols differed.Cox used combined training and validation sets (n = 227), whereas the neural model used training data (n = 152); the comparison does not isolate model class.Source: §5 · p. 10
- 0.652 using numeric variables exceeded mutation-extended Cox results with BioClinical ModernBERT embeddings (0.594) and CONCH embeddings (0.538).Encoder and reader differences were reported descriptively, not as evidence of their underlying causes.Source: §5 · p. 10
- 0.603 ± 0.030 was the image-only concept-guided attention model's mean C-index, the highest among single-modality neural configurations.The image-only attention and mean-pooled image Cox pipelines differed in token selection, aggregation, survival model, and fitting protocol, so their performance difference cannot be attributed to aggregation alone.Source: §5 · p. 10
- Co-attention maps over WSI tokens were generated per variable, but were not treated as validated explanations.Across-repetition standard deviations were 0.025 for fusion and 0.042 for clinical-only Cox; with five repetitions, this difference is descriptive rather than evidence of greater stability.Source: §5 · pp. 10–11
- 0.652 ± 0.046 was the highest mean C-index overall for mutation-extended numeric Cox regression; multimodal co-attention led neural models at 0.619 ± 0.025.The fused neural model exceeded image-only attention (0.603 ± 0.030) but remained below clinical-only Cox (0.649 ± 0.042).Source: §5 · p. 10, Table 2 · p. 11
6 Conclusion
The paper introduces a patient-level multimodal dataset of resected PDAC and reports initial survival benchmarks across clinical, molecular, histological, and fused approaches. These benchmarks provide a basis for future pancreas-specific modeling and cross-cohort validation.
6 Conclusion
- The dataset links digitised H&E whole-slide images with clinicopathological variables, targeted mutation data, and long-term overall-survival outcomes at the patient level.The common outcome definition supports evaluation of clinical, molecular, histological, and multimodal approaches.Source: §6 · p. 11
- 0.652 ± 0.046 was the highest mean concordance overall, while multimodal co-attention achieved the highest neural concordance at 0.619 ± 0.025.The image-only attention model achieved 0.603 ± 0.030; these are initial benchmarks, not evidence of an optimal strategy.Source: §6 · p. 11
- The dataset and benchmarks provide a foundation for future pancreas-specific foundation models, patient-level WSI aggregation, multimodal survival modeling, and cross-cohort validation.Source: §6 · p. 11
B.1 Targeted sequencing data availability
Targeted sequencing results were available for about half of the patients with whole-slide image features; missingness was at the patient level.
B.1 Targeted sequencing data availability
- Sequencing was available for 154 of 302 patients (51.0%) and unavailable for 148 (49.0%); all 102 panel genes were assessed for each sequenced patient.Thus, the reported missingness was patient-level, with no gene-level gaps among patients with panel results.Source: §B.1 · p. 18
B.2 Mutation profiles of sequenced patients
The 154 sequenced patients had pathogenic or likely pathogenic variants summarized across 102 genes; KRAS and TP53 were most frequently mutated.
B.2 Mutation profiles of sequenced patients
- KRAS mutations affected 133 (86.4%) patients and TP53 mutations affected 78 (50.6%) in the sequenced WSI cohort.Mutations in other panel genes were uncommon, and patients had zero to four mutated genes.Source: §B.2 · p. 18, Figure 2 · p. 20