BreastXAgent: A lightweight agentic explainability framework for breast density classification
Deep learning models for breast density classification typically return only a label and confidence score. This paper presents BreastXAgent, a lightweight agentic framework that combines a pretrained classifier, Grad-CAM visual explanation, LangGraph-based routing, LLM report generation, and rule-based guardrails to produce controlled, faithful, non-diagnostic reports. On 160 mammography images, uncontrolled LLM reporting produced 75% Grad-CAM overclaims and 62.5% clinical overclaims. Strict prompting and LangGraph routing improved safety but did not eliminate violations. The full BreastXAgent achieved 100% safety pass rate, zero overclaims, and highest utility score. Results demonstrate that agentic guardrails enable safe, structured explanation generation without modifying the underlying classifier.