Assessing Pancreatic Ductal Adenocarcinoma Vascular Invasion: the PDACVI Benchmark
| Authors | M. Riera-Marín et al. |
| Year | 2026 |
| HF Upvotes | 3 |
| arXiv | 2604.27582 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
Surgical resection remains the only potentially curative treatment for pancreatic ductal adenocarcinoma (PDAC), and eligibility depends on accurate assessment of vascular invasion (VI), i.e., tumor extension into adjacent critical vessels. Despite its importance for preoperative staging and surgical planning, computational VI assessment remains underexplored. Two major challenges are the lack of public datasets and the diagnostic ambiguity at the tumor-vessel interface, which leads to substantial inter-rater variability even among expert radiologists. To address these limitations, we introduce the CURVAS-PDACVI Dataset and Challenge, an open benchmark for uncertainty-aware AI in PDAC staging based on a densely annotated dataset with five independent expert annotations per scan. We also propose a multi-metric evaluation framework that extends beyond spatial overlap to include probabilistic calibration and VI assessment. Evaluation of six state-of-the-art methods shows that strong global volumetric overlap does not necessarily translate into reliable performance at clinically critical tumor-vessel interfaces. In particular, methods optimized for binary segmentation perform competitively on average overlap metrics, but often degrade in high-complexity cases with low expert consensus, either collapsing in volume or overextending at uncertain boundaries. In contrast, methods that model inter-rater disagreement produce better calibrated probabilistic maps and show greater robustness in these ambiguous cases. The benchmark highlights the limitations of volumetric accuracy as a proxy for localized surgical utility, motivating uncertainty-aware probabilistic models for preoperative decision-making.
Engineering Breakdown
Plain English
This paper addresses a critical gap in pancreatic cancer diagnosis by introducing CURVAS-PDACVI, the first public benchmark dataset for detecting vascular invasion (tumor extension into blood vessels) in pancreatic ductal adenocarcinoma using AI. The core problem is that vascular invasion assessment is essential for determining surgical eligibility and treatment planning, but it's subjective—even expert radiologists disagree significantly on whether tumors have invaded adjacent vessels. The authors solved this by creating a densely annotated dataset with five independent expert annotations per scan, allowing them to study uncertainty and inter-rater variability as a core challenge rather than noise. This enables the research community to develop AI models that are uncertainty-aware, meaning they can quantify confidence in predictions rather than just outputting binary decisions.
Core Technical Contribution
The primary innovation is framing vascular invasion assessment as an uncertainty quantification problem rather than a standard classification task. Instead of treating disagreement between expert radiologists as noise to average out, the authors preserve and leverage this disagreement by requiring five independent annotations per scan, enabling models to learn the inherent ambiguity at tumor-vessel boundaries. The CURVAS-PDACVI dataset itself is a significant contribution—the first public, densely annotated benchmark in this domain addresses the documented lack of open data that has kept computational VI assessment underexplored despite its clinical importance. This shifts the paradigm from deterministic prediction (tumor invades vessel: yes/no) to probabilistic modeling (likelihood of invasion with calibrated uncertainty bounds), which is more clinically actionable because it surfaces ambiguous cases for human review.
How It Works
The technical approach centers on using the multiple expert annotations to model and predict uncertainty in vascular invasion assessment. For each patient scan (likely multi-phase CT or MRI data), radiologists independently annotate the tumor-vessel interface, marking regions where invasion is present, absent, or ambiguous. The dataset preserves all five annotations rather than creating a single ground truth label, enabling downstream models to learn the distribution of expert opinions and estimate prediction uncertainty through methods like Bayesian deep learning, ensemble approaches, or direct uncertainty regression. The input is volumetric imaging data (3D scans showing tumor and surrounding vasculature); the model processes this through a medical imaging backbone (likely a 3D U-Net or similar segmentation architecture); and the output is both a probabilistic prediction of vascular invasion and an uncertainty estimate reflecting how confident the model should be. This multi-annotator framework allows researchers to evaluate models not just on accuracy but on calibration—whether predicted confidence matches actual correctness—which is crucial for clinical decision support where false confidence is dangerous.
Production Impact
In a production surgical planning pipeline, this uncertainty-aware approach would fundamentally change how AI assists radiologists in assessing operability. Rather than a binary recommendation ('patient is/isn't a surgical candidate'), the system would flag cases with high prediction uncertainty, routing them for mandatory expert review while handling high-confidence cases more efficiently. This reduces false negatives (incorrectly clearing a patient for surgery when vascular invasion is present) by surfacing ambiguous cases, improving patient safety. The multi-annotator training approach also makes the system more robust—it learns that vascular invasion is genuinely difficult at certain anatomical boundaries, rather than overfitting to one radiologist's interpretation style. However, the trade-offs are significant: you need densely annotated data (five experts per scan is expensive to acquire), inference latency increases slightly when computing uncertainty estimates, and the model requires careful calibration to ensure uncertainty scores are meaningful in the specific clinical context where it's deployed. Integration requires retraining on your institution's imaging protocols, as MRI vs. CT and reconstruction parameters affect tumor-vessel visibility.
Limitations and When Not to Use This
The paper focuses on vascular invasion detection but doesn't address other critical aspects of PDAC staging (tumor size, metastasis, margin status), so it's a narrow-scope solution within surgical planning. The dense annotation requirement (five experts per case) is resource-intensive and may not be achievable in many medical centers, limiting reproducibility and further dataset expansion—it also introduces selection bias if only certain cases are heavily annotated. The dataset appears to be from a single institution or small number of centers, which likely means the learned uncertainty patterns may not generalize to different imaging equipment, protocols, or patient populations (different ethnic groups, comorbidities, etc.). The abstract doesn't specify which imaging modalities or phases are included, so it's unclear whether the model handles multimodal data or requires specific sequences; in production, missing data (e.g., arterial phase unavailable) could degrade performance. Finally, the paper doesn't appear to address temporal data—surgical outcomes validation (did the model's uncertainty predictions correlate with actual intraoperative findings?) is not mentioned, so clinical utility remains unproven despite the methodological soundness.
Research Context
This work builds on decades of clinical radiology research establishing vascular invasion as a key prognostic factor in PDAC, but brings modern uncertainty quantification and deep learning to a problem previously solved only through subjective expert interpretation. It extends the broader trend in medical AI of treating inter-rater variability not as noise but as informative signal—similar recent work includes uncertainty quantification in pathology and multi-reader studies in radiology benchmarks. The paper directly addresses a documented gap in public datasets for pancreatic cancer AI (prior work relied on proprietary, single-institution datasets), positioning CURVAS-PDACVI as infrastructure for the research community, akin to how datasets like ImageNet or MICCAI challenges accelerated progress in computer vision. This opens a new research direction: shifting from 'can we match expert accuracy?' to 'can we quantify when experts disagree and learn from that disagreement?', which has applications across medical imaging tasks where true ground truth is inherently subjective.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::