Deep learning CT model for stratified diagnosis of pancreatic cystic neoplasms: multicenter development, validation, and real-world clinical impact - Report - MDSpire
Advertisement
Deep learning CT model for stratified diagnosis of pancreatic cystic neoplasms: multicenter development, validation, and real-world clinical impact
AI-Enhanced CT Imaging Model for Improved Diagnosis of Pancreatic Cystic Neoplasms
Overview
This multicenter study developed and validated PCN-AI, an AI-powered CT imaging model that significantly improves diagnostic accuracy and efficiency for pancreatic cystic neoplasms (PCN). The model demonstrated robust performance across multiple classification tasks and real-world clinical settings, enhancing early detection and reducing radiologist workload.
Background
Pancreatic cystic neoplasms are important precursors for early pancreatic cancer detection, but current diagnostic methods suffer from variability and limited accuracy. Conventional imaging criteria have limitations, and invasive procedures carry risks and false negatives. Artificial intelligence offers potential to improve non-invasive diagnosis by providing quantitative, interpretable imaging biomarkers and hierarchical classification of PCN subtypes, yet prior models lacked multicenter validation and clinical integration.
Data Highlights
Metric
Value
Number of patients included
1835
Number of quantitative features extracted
63
Diagnostic accuracy improvement (AUC) with AI assistance
From 0.786 to 0.845 (p < 0.05)
Reduction in interpretation time
23.7% (5.28 to 4.03 minutes per case)
Radiologist acceptance of AI recommendations
87.14%
Clinical workload reduction in real-world cohort
39.3%
Diagnostic benefit to patients in real-world cohort
45.45% (5/11 patients with missed malignant PCN correctly identified)
Model AUC range across classification tasks
0.845–0.988
Key Findings
PCN-AI extracted 63 quantitative imaging features from contrast-enhanced CT scans to classify PCN subtypes hierarchically.
AI assistance significantly improved radiologists’ diagnostic accuracy (AUC increase from 0.786 to 0.845) and reduced interpretation time by nearly 24%.
Radiologists accepted AI recommendations in over 87% of cases, indicating high clinical trust and usability.
In a prospective real-world cohort, PCN-AI outperformed traditional double-reading, correctly identifying nearly half of missed malignant cases and reducing workload by 39.3%.
The model demonstrated robust performance across four classification tasks, distinguishing mucinous vs. non-mucinous PCN, precancerous vs. malignant lesions, and differentiating key subtypes (IPMN, MCN, SCN, SPN) with AUCs up to 0.988.
Clinical Implications
PCN-AI can be integrated into radiologist workflows to enhance early detection and precise classification of pancreatic cystic neoplasms, facilitating timely intervention for malignant cases. Its ability to reduce interpretation time and workload may improve clinical efficiency and consistency, addressing current diagnostic variability and limitations of invasive procedures.
Conclusion
The PCN-AI model offers a validated, interpretable, and clinically impactful tool for improving the diagnosis and management of pancreatic cystic neoplasms across multiple centers. Its adoption has the potential to enhance early pancreatic cancer detection and optimize patient outcomes.
References
Study Authors/2024 -- AI-Enhanced CT Imaging Model for Improved Diagnosis of Pancreatic Cystic Neoplasms
