StrokeChat: A Locally Deployable Conversational Interface for Stroke-Imaging AI Models

A growing number of AI models can detect, segment, and characterize stroke-related findings on brain imaging. However, many remain difficult for clinicians to access because they are distributed as model weights, code repositories, and technical pipelines. This access gap may be particularly important in smaller hospitals and resource-limited settings where immediate neuroradiology expertise is not continuously available. We developed \emph{StrokeChat}, a clinician-facing platform that integrates stroke lesion detection and segmentation, ASPECTS estimation, cerebrovascular segmentation, and medical vision-language conversation within a single interface. StrokeChat is designed to run fully locally, as a downloadable application that can be installed on institutional hardware so that patient imaging never leaves the site; a web version is also available for convenient testing and demonstration. In a prospective evaluation, ten physicians (five neurologists and five radiologists) independently reviewed five de-identified CT/CTA cases each, yielding 50 case-level evaluations. \ph{insert final usability, grounding, clinical-reasonableness, and safety results}. StrokeChat illustrates a practical, locally deployable approach for translating specialized stroke-imaging AI models into an accessible, conversational decision-support environment that keeps imaging data on-site.