Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Jul 24·5 min read
Spatially Grounded Concept Bottleneck Models for Trustworthy Breast Ultrasound Diagnosis
TL;DR
A Spatially Grounded Concept Bottleneck Model (SG-CBM) was developed to enhance the trustworthiness of breast ultrasound diagnoses by using weak supervision from lesion delineations. This model significantly improved diagnostic accuracy and spatial alignment of concept evidence, addressing issues of interpretability in medical imaging.
✦ Why It Matters
Implement data-quality-aware supervision in AI models to enhance diagnostic trustworthiness in healthcare applications.
Key Takeaways
- Utilize weak supervision from lesion delineations for model training.
- Focus on spatially relevant concept evidence in medical imaging.
- Conduct quality stress tests to evaluate supervision impact.
- Enhance interpretability of AI models in healthcare diagnostics.
How It Works
The SG-CBM uses weak supervision from coarse lesion delineations to guide the model in identifying anatomically relevant concepts. It defines two key regions from each lesion mask: an in-lesion area for morphology-related concepts and a posterior acoustic band for additional diagnostic features.
By employing a grouped spatial grounding objective, the model ensures that the concept evidence aligns spatially with the actual lesions, improving interpretability and trustworthiness.