Journal: NPJ digital medicine
This study describes UPATHLN, a pan-cancer AI system for detecting lymph node metastases on pathology slides, with a specific focus on safety and reduction of overconfident errors in rare histologic variants.
Key points:
- • Clinical problem: Histologic heterogeneity across many primary tumor types leads existing AI tools to make confident but incorrect calls on rare or unfamiliar variants, risking missed nodal metastases.
- • Methodology:
- Uses a pathology foundation model–based encoder to analyze lymph node histology.
- Couples this with a decoupled uncertainty estimation module that flags cases where the model is likely to be wrong—especially potential false negatives—for mandatory pathologist review.
- Trained and validated on a large multicenter dataset of 26,229 lymph nodes from 14 different primary tumor origins.
- • Performance:
- Internal validation AUC: 0.986 for metastasis detection.
- The uncertainty module identified all potentially missed metastases, achieving 100% conditional sensitivity in both development and independent test cohorts, including tumors from seven primary sites not seen during training.
- • Workflow impact:
- For lymph nodes classified as negative, the uncertainty mechanism reduced the pathologist review workload by 73.2%, while still capturing all metastatic nodes via flagged “uncertain” cases.
- • Implications for oncology practice:
- Demonstrates that explicit modeling of uncertainty can transform AI from a simple classifier into a safety-critical triage and decision-support system.
- Offers a potential path to reliable, workload-efficient nodal assessment across multiple tumor types, with particular relevance to minimizing missed micrometastases and uncommon histologies.