MammoWise: Multi-Model Local RAG Pipeline for Mammography Report Generation

A research team including Raiyan Jahangir and Nafiz Imtiaz Khan has introduced MammoWise, an innovative local AI pipeline designed to automate mammography report generation while addressing critical limitations of existing cloud-based systems. Published as an arXiv preprint, MammoWise transforms open-source Vision Language Models (VLMs) into specialized tools for creating the structured narrative reports, BI-RADS assessments, and breast density classifications that radiologists must produce. The system's key innovation is its completely local, privacy-preserving architecture that supports any VLM hosted via Ollama, enabling healthcare institutions to deploy AI assistance without sending sensitive patient data to external servers. This addresses significant concerns about data security, reproducibility, and adaptability that plague many current medical AI implementations.

The technical framework combines multiple approaches: it enables zero-shot, few-shot, and Chain-of-Thought prompting, and incorporates optional multimodal Retrieval-Augmented Generation (RAG) using a ChromaDB vector database to provide case-specific context. The researchers evaluated models including MedGemma, LLaVA-Med, and Qwen2.5-VL on the VinDr-Mammo and DMID datasets, measuring report quality via BERTScore and ROUGE-L metrics alongside classification accuracy. Parameter-efficient fine-tuning using QLoRA significantly boosted MedGemma's performance, achieving 0.8840 accuracy for breast density classification and 0.7545 for BI-RADS assessment while maintaining strong report generation capabilities. MammoWise represents a practical, extensible framework that could accelerate clinical workflow by automating documentation while giving medical professionals full control over their AI tools and data.