Uncertainty-Aware Deep Learning Model for Pancreatic Cancer Detection

Overview

Pancreatic cancer remains one of the deadliest cancers worldwide, as pancreatic tumors are small and often obscured by surrounding anatomical structures. Computed tomography (CT) is the primary imaging modality used for pancreatic tumor detection; however, diagnostic uncertainty exceeds 50%, and wait times for definitive diagnosis can exceed 30-60 days despite the rapidly metastasizing nature of the disease. While computational models have shown promise in medical image analysis, clinical implementation remains limited due to the absence of complete diagnostic pipelines, limited interpretability, and insufficient understanding of model performance under real-world conditions. This study develops a deployable AI-based pipeline for pancreatic tumor segmentation, introduces uncertainty estimation to improve model transparency, and investigates the relationship between imaging acquisition parameters and the reliability of AI-based tumor detection. Using an nnU-Net-based deep learning model and 9,900 CT images, pancreatic tumors were segmented from CT scans. Predictive uncertainty was quantified using SoftMax-derived entropy, allowing assessment of model confidence alongside segmentation results. A complete analysis pipeline was constructed, from CT input to segmentation and uncertainty map outputs. The relationship between CT slice thickness and predictive uncertainty was analyzed to determine how imaging parameters influence AI reliability. Results indicate that moderate slice thickness values (approximately 2.5-3 mm) are associated with lower model uncertainty, whereas extremely thin or thick slices increase uncertainty. These findings demonstrate that CT acquisition parameters can significantly influence the reliability of computational tumor detection systems and help bridge the gap between experimental AI models and clinically deployable diagnostic tools with real-world impact.

Competition history

• CSEF 2026 Medicine & Physiology (Track 2) (Senior Division) · Entry S-21-29

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Biomedical and Health Sciences projects
• CSEF Biomedical and Health Sciences projects
• Every CSEF 2026 project