High-sensitivity pan-cancer AI assessment of lymph node metastasis via uncertainty quantification

Area of Science:

• Artificial Intelligence in Pathology
• Computational Oncology
• Diagnostic AI

Background:

• Histological heterogeneity in primary tumors complicates lymph node metastasis assessment.
• Existing AI systems struggle with rare cancer variants, leading to overconfident errors and missed diagnoses.

Purpose of the Study:

• To develop and validate UPATHLN, a unified diagnostic platform for accurate pan-cancer lymph node metastasis assessment.
• To integrate a pathology foundation model with uncertainty estimation to enhance AI diagnostic reliability.

Main Methods:

• Developed UPATHLN by combining a pathology foundation model encoder with a decoupled uncertainty estimation mechanism.
• Validated the system on a large-scale multicenter dataset of 26,229 lymph nodes from 14 primary origins.
• Evaluated the uncertainty module's performance in flagging potential false-negative predictions.

Main Results:

• UPATHLN achieved an Area Under the Curve (AUC) of 0.986 in internal validation.
• The uncertainty module successfully intercepted all missed diagnoses, achieving 100% conditional sensitivity across development and test cohorts, including unseen origins.
• The system reduced the review burden for negative lymph nodes by 73.2%.

Conclusions:

• UPATHLN establishes a new benchmark for safety-critical AI in diagnostics.
• Explicitly modeling uncertainty is crucial for reliable and workload-efficient AI diagnostics at a pan-cancer scale.
• The platform demonstrates the potential for AI to significantly improve diagnostic accuracy and efficiency in pathology.