Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·May 25·8 min read
PathNavigate: A Training-Free Pathology Agent with Surprise-Guided Scan and Shared Slide Memory for Whole-Slide Image VQA
TL;DR
Pathologists need to answer clinical questions about gigapixel microscope slides (whole-slide images) but can only inspect a tiny fraction due to computational limits. PathNavigate is a training-free agent that first scans slides at low magnification to identify abnormal regions, then searches within those regions for question-relevant evidence at high magnification.
✦ Why It Matters
Engineers can deploy PathNavigate on new pathology tasks without retraining, reducing development time and enabling practical clinical decision support.
Key Takeaways
- Use surprise fields to pre-filter abnormal regions before question matching, avoiding missed morphology.
- Deploy frozen models without task-specific retraining to reduce engineering overhead and enable rapid adaptation.
- Implement online memory modules to maintain slide-level context across low and high magnification scans.
- Combine low-magnification scanning with high-magnification evidence extraction to respect computational budgets on gigapixel images.
Full Summary
Whole-slide image visual question answering (WSI-VQA) requires systems to locate sparse diagnostic evidence within gigapixel pathology images under strict computational budgets. Existing approaches either couple navigation to task-specific training (limiting reusability) or use question-first designs that miss morphological abnormalities not mentioned in the query.
PathNavigate introduces a scan-search-readout routine: it first performs low-magnification scanning using a shared online memory module over frozen pathology features to generate a surprise field marking abnormal regions, then applies question-conditioned relevance matching only within that pool to select high-magnification targets, and finally extracts evidence using frozen perceptor-adjudicator components. The method keeps all core models frozen, avoiding retraining costs while improving answer accuracy on WSI-VQA and SlideBench-BCNB benchmarks.
Results show more interpretable evidence-selection trajectories and better handling of morphology not explicitly named in clinical queries.