Agentic Multimodal Architectures for Medical Imaging: Orchestration, Deterministic Fairness Probing, and Governance

Large Multimodal Models (LMMs) are promising for medical-image interpretation but struggle with multi-step diagnostic workflows over image series, longitudinal tracking across imaging sessions, and demographic shortcut learning. This Work-in-Progress presents an open-source agentic framework that addresses these gaps with a deterministic, auditable design rather than a single monolithic prompt. A LangGraph state-graph orchestrates a Fan-Out/Fan-In pipeline — per-image analysis, per-series aggregation, and cross-session evolution — over a Google Gemini 2.5 backend. Two governance components are engineered as first-class, test-covered artefacts: a deterministic, explainable allocative-harm probe that flags demographic-anchored justifications in generated text as a regression guard, and a token-level cost meter with an enforceable spend cap. The implementation is verified by 124 automated tests and cross-walked against the EU AI Act and the NIST AI Risk Management Framework. A live end-to-end run on the public Gemini API reports measured token usage, cost, and latency. The system is offered as a reproducible reference for governance-aware clinical AI tooling; the full source, test suite, and compliance matrices are released under GPL-3.0.