[Paper] Early and Prediagnostic Detection of Pancreatic Cancer from Computed Tomography

Source: arXiv - 2601.22134v1

Overview

Pancreatic ductal adenocarcinoma (PDAC) is usually diagnosed at an advanced stage, making curative treatment almost impossible. This paper introduces ePAI, an AI‑driven system that scans routine abdominal CT images and flags lesions as small as a few millimeters, enabling detection months before clinicians would normally notice them. The authors demonstrate that ePAI can outperform radiologists on both internal and multi‑center external datasets, suggesting a realistic path toward earlier, life‑saving interventions.

Key Contributions

• End‑to‑end detection pipeline (segmentation + classification) that localizes PDAC lesions ≤ 2 mm on standard contrast‑enhanced CT scans.
• Large‑scale training and validation: 1,598 patients for training, 1,009 internal test cases, and an external cohort of 7,158 patients from six hospitals.
• State‑of‑the‑art performance: AUC 0.939–0.999 (internal) and 0.918–0.945 (external); sensitivity ≈ 95 % (internal) and 91 % (external) with high specificity.
• Lead‑time advantage: correctly identified 75/159 pre‑diagnostic scans, giving a median of 347 days of earlier detection.
• Multi‑reader study: ePAI achieved a 50 % higher sensitivity than a panel of 30 board‑certified radiologists while keeping specificity comparable (≈ 95 %).
• Open‑source‑ready architecture: the paper details model components (3‑D CNN backbone, attention‑guided localization) and training recipes that can be reproduced with publicly available deep‑learning frameworks.

Methodology

1. Data Curation – All CT scans were retrospectively collected, de‑identified, and annotated by expert radiologists for tumor presence, size, and location.
2. Pre‑processing – Scans were resampled to a uniform voxel spacing, intensity‑normalized, and pancreas‑specific region‑of‑interest (ROI) cropping was applied to reduce background noise.
3. Model Architecture –
   • A 3‑D convolutional backbone (ResNet‑3D variant) extracts volumetric features.
   • An attention module highlights suspicious regions, feeding a segmentation head that produces a heatmap of potential lesions.
   • A classification head aggregates the heatmap to output a binary “cancer / no‑cancer” score.
4. Training Strategy – A combination of cross‑entropy loss (for classification) and Dice loss (for segmentation) was optimized jointly. Hard‑negative mining and data augmentation (rotation, scaling, intensity jitter) helped the model generalize to different scanners and protocols.
5. Evaluation – ROC curves, sensitivity/specificity at clinically relevant operating points, and lesion‑localization accuracy (distance between predicted and ground‑truth centroids) were measured. A separate reader study compared AI outputs against radiologists’ interpretations on the same cases.

Results & Findings

• Lead‑time: In the pre‑diagnostic cohort, ePAI flagged tumors an average of ≈ 11 months before the radiologist’s first report.
• Reader Study: Radiologists achieved ~33 % sensitivity; ePAI raised this to ~50 % higher while keeping specificity at 95.4 %.
• Localization Accuracy: Median centroid error < 4 mm, well within the size of early lesions, confirming that the system can guide follow‑up imaging or biopsy.

Practical Implications

• Clinical Decision Support – ePAI can be integrated into PACS/RIS workflows as a “second reader,” automatically flagging high‑risk slices for radiologist review, reducing missed early lesions.
• Population Screening – Hospitals that already acquire abdominal CTs for unrelated reasons (e.g., trauma, abdominal pain) could retroactively run ePAI on archived scans, creating a low‑cost opportunistic screening program.
• Research & Trial Enrichment – Early‑stage PDAC patients identified by ePAI could be enrolled in neoadjuvant or targeted‑therapy trials, accelerating drug development pipelines.
• Software Development – The modular architecture (3‑D CNN + attention) can be repurposed for other small‑lesion detection tasks (e.g., early liver metastases, lung nodules), offering a reusable codebase for AI‑enabled radiology tools.
• Regulatory Pathway – Demonstrated performance across multiple centers and a reader study aligns with FDA’s “Software as a Medical Device” (SaMD) evidentiary requirements, smoothing the path toward clinical clearance.

Limitations & Future Work

• Dataset Bias – Training data came from a single institution; despite external validation, subtle scanner‑specific artifacts could affect generalization to low‑resource settings.
• Retrospective Design – The study used already‑collected scans; prospective trials are needed to confirm real‑time impact on patient outcomes and workflow efficiency.
• Interpretability – While attention maps provide visual cues, deeper explainability (e.g., radiomics‑based reasoning) could increase clinician trust.
• Integration Challenges – Deploying a 3‑D CNN at scale requires GPU resources; future work should explore model compression or edge‑computing solutions.
• Broader Clinical Context – Combining ePAI with serum biomarkers (CA‑19‑9, circulating tumor DNA) could further improve early‑detection specificity and reduce false positives.

Overall, ePAI showcases how modern computer‑vision techniques can move from research labs into tangible tools that give developers and healthcare providers a concrete lever to fight one of the deadliest cancers.

Authors

• Wenxuan Li
• Pedro R. A. S. Bassi
• Lizhou Wu
• Xinze Zhou
• Yuxuan Zhao
• Qi Chen
• Szymon Plotka
• Tianyu Lin
• Zheren Zhu
• Marisa Martin
• Justin Caskey
• Shanshan Jiang
• Xiaoxi Chen
• Jaroslaw B. Ćwikla
• Artur Sankowski
• Yaping Wu
• Sergio Decherchi
• Andrea Cavalli
• Chandana Lall
• Cristian Tomasetti
• Yaxing Guo
• Xuan Yu
• Yuqing Cai
• Hualin Qiao
• Jie Bao
• Chenhan Hu
• Ximing Wang
• Arkadiusz Sitek
• Kai Ding
• Heng Li
• Meiyun Wang
• Dexin Yu
• Guang Zhang
• Yang Yang
• Kang Wang
• Alan L. Yuille
• Zongwei Zhou

Paper Information

• arXiv ID: 2601.22134v1
• Categories: cs.CV
• Published: January 29, 2026
• PDF: Download PDF