A standards-based infrastructure for distributed AI inference in medical imaging

Background: Integrating artificial intelligence (AI) into clinical workflows requires secure, interoperable infrastructures that connect heterogeneous clinical data systems while respecting governance and regulatory constraints. In medical imaging, PACS vendors implement the DICOMweb standard with significant variation, creating interoperability challenges for federated AI inference. Methods: This work presents a standards-based DICOMweb adapter that mediates between heterogeneous clinical PACS implementations and federated AI services through protocol translation between DICOMweb and DIMSE, configuration-driven normalization of vendor-specific deviations, and explicit extension points for governance services such as de-identification. The approach was evaluated through clinical deployment at a university hospital and a reproducible multi-PACS simulation with two independent Orthanc backends, with a comparative benchmark against Orthanc's DICOMweb plugin for context. Results: The adapter showed consistent performance across heterogeneous backends (median QIDO-RS latency 46–50 ms, 100% success rate). The benchmark showed that, in its default configuration, Orthanc does not proxy DICOMweb requests to remote DIMSE-based PACS and returns HTTP 404 for all WADO-RS retrievals requiring the multipart content type expected by downstream AI pipelines. The adapter served all requests successfully. Discussion: Existing open-source tools are not designed for the mediation role required in federated AI workflows; a dedicated, standards-based mediation layer is necessary to bridge the gap between heterogeneous clinical PACS and distributed AI services.