MIRA: An Edge-Cloud Multimodal LLM Framework for DESH Assessment in IoT-Enabled MRI Workflows

Abstract

Connected MRI scanners operate as networked medical devices within DICOM-based healthcare Internet of Things (IoT) workflows, but conventional connectivity primarily moves images to a Picture Archiving and Communication System (PACS) rather than adding task-specific reasoning near acquisition. Building on DESH-App 2.0, a preceding machine-learning contribution accepted for oral presentation at SIN Firenze 2026, this paper specifies MIRA (Multimodal Imaging Reasoning and Assessment), a vendor-neutral edge-cloud module for assessing disproportionately enlarged subarachnoid spaces (DESH). A hospital-edge gateway validates and de-identifies 3D T1-weighted DICOM studies, derives ventricular, high-convexity, and Sylvian-fissure masks with SegFormer3D, and converts each volume into a compact evidence package comprising 12 representative slices with overlays, 17 global, zonal, and ratio features, and quality-control tokens. LLaVA-Med-7B and Lingshu-7B are adapted through 4-bit QLoRA to return constrained JSON containing class, confidence, named anatomical evidence, and an abstention or quality warning. The edge-cloud partition keeps raw volumes inside the institutional network; only the compact evidence package may reach an authenticated inference service, while structured outputs return to the PACS workstation. The evaluation protocol jointly covers diagnostic discrimination, calibration, evidence agreement, unsupported claims, abstention, latency, transmitted bytes, memory demand, link-loss recovery, and multi-device interoperability. MIRA is designed to add auditable multimodal reasoning to connected MRI workflows without modifying safety-critical scanner controls.