Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 24·5 min read
TRACE: Training-time Report-guided and Clinically Ordered Concept Editing
TL;DR
TRACE is a novel framework that enhances breast ultrasound diagnosis by utilizing structured radiology reports for concept editing, enabling image-only diagnosis at test time. It demonstrates superior performance and robustness across multiple datasets compared to existing methods.
✦ Why It Matters
Implement TRACE to enhance the interpretability and robustness of your breast ultrasound diagnostic models today.
Key Takeaways
- Utilize TRACE for improved interpretability in medical imaging.
- Incorporate structured radiology reports for concept supervision.
- Apply Strategic Concept Missing Training for better annotation handling.
- Leverage BUSC benchmark for enhanced model evaluation.
How It Works
TRACE employs a teacher-guided editing mechanism to refine image-derived concepts, organizing them within a malignancy-aware ordered concept space. This allows the model to learn from structured reports while still performing image-only diagnosis at test time.