FERMED: A Description-First Framework for Reviewable Medical Image Interpretation

Medical vision-language systems often combine image description, disease interpretation and report generation in a single response. This makes it difficult to identify whether an error arose from visual observation, clinical inference or communication. FERMED proposes a description-first workflow that separates these responsibilities through an explicit, inspectable interface. A visual module records image quality, observable findings and unavailable evidence; an interpretation module uses those observations to formulate bounded clinical hypotheses; and a consistency module checks whether each statement is supported by the recorded findings. A clinician retains access to the original image and decides whether to accept, revise or reject the report. Ophthalmic imaging provides a motivating application because apparently similar findings can support different interpretations and some diagnoses require information absent from a single image. The contribution is a model-agnostic architecture, a structured output contract and an evaluation protocol that measures observation fidelity separately from downstream decision performance. The protocol includes patient-level separation, external validation, description perturbations, matched-compute baselines, calibration and selective prediction. This paper does not report a trained FERMED model, patient cohort, diagnostic accuracy result or clinical validation. It establishes testable design hypotheses and a reproducible framework for evaluating them before any clinical use.