The Standard

 Overall Winner

The  Integrated Medical Management and Educational Gateway (IMMEG) Venous Management System (I-VMS) is an AI-enabled, standards-based platform that projects a vectorized atlas of the deep thoracic venous system onto routine chest radiographs. Using deep-learning landmark detection and HL7 FHIR®/DICOM interoperability, the system lets clinicians visualize the catheter trajectory and tip position in patient-specific anatomy, record planned versus actual placement, and build a reusable, longitudinal venous access record across organizations. The project was developed for Vanguard with support from Xzyos.ai. 

Clinical Problem and Context 

Central venous access is essential for chemotherapy, parenteral nutrition, dialysis and critical care, yet malposition and related complications—venous injury, thrombosis, infection, and device dysfunction—remain common and costly. Post-procedure assessment usually relies on plain chest X-rays, which do not directly visualize venous structures. As a result, clinicians infer anatomy indirectly; documentation is inconsistent; and comparing procedures over time is difficult. There is no consolidated, spatially normalized record of a patient’s venous history to guide future decisions.

Core Innovation

 I-VMS predicts anatomical landmarks (e.g., carina, first thoracic vertebra T1, lateral edge of the right rib) on a radiograph with a modified DenseNet121 model implemented in MONAI. These coordinates establish a patient-specific basis for an affine transformation that overlays a standardized, vector-based venous atlas onto the image. Clinicians can accept or adjust landmarks and annotate intended and actual entry and tip positions. Because annotations are stored in a normalized coordinate space, results are comparable across encounters and over time, enabling longitudinal analysis and population-level learning.

Architecture

The web application uses Angular with NgRx for a predictable, testable, unidirectional state flow (components dispatch actions; reducers update state; effects handle asynchronous work). The serverless backend on AWS employs API Gateway, Cognito, and Lambda for secure, stateless logic; DynamoDB for structured data; OpenSearch for indexed queries and analytics; S3 for encrypted storage; and CloudWatch for monitoring. A medical integration layer connects a FHIR server (clinical and imaging metadata) and a DICOM server (image retrieval via DICOMweb/WADO-RS). This design scales automatically, preserves local control of images, and keeps the system standards-compliant.

Model and Training

DenseNet121’s classification head is replaced with a coordinate-regression head that outputs (x, y) pairs for each landmark. Training used 180 expert-annotated chest radiographs rescaled to 512×512 and normalized, with an 80/20 train-test split. Optimization employed AdamW with ReduceLROnPlateau scheduling, early stopping, and mean-squared-error (“LandmarkLoss”). Albumentations applied landmark-preserving augmentations—small rotations, shears, brightness/contrast adjustments, and Gaussian noise—to increase diversity and reduce overfitting. On an RTX A5000 laptop GPU, validation error plateaued at ~2.7% for this prototype; further improvement toward <1% will require larger datasets and more compute.

Standards-driven Workflow

FHIR’s ImagingStudy indexes DICOM series and references a FHIR Endpoint that describes how to retrieve images from the organization’s DICOMweb service. I-VMS queries FHIR for ImagingStudy and Endpoint, fetches DICOM images from source archives, invokes the AI model for landmarks, projects the venous atlas, and renders an interactive view for clinician confirmation and procedural annotation. The resulting record is traceable, spatially normalized, and reusable for longitudinal tracking and quality improvement.

 Figure 1: Sequence Flow - Initiate call for ImagingStudy through Image Rendering and Atlas Projection 

The figure below depicts the process of enriching medical images with AI-generated and physician-provided annotations, ensuring accuracy, traceability, and alignment with standardized anatomical references. The workflow begins with secure retrieval of the patient’s DICOM image from the imaging server. The AI model then analyzes the image to identify key anatomical landmarks, which are projected onto the image for review. Physicians can accept or adjust these landmarks, after which the confirmed points are used to project the Atlas coordinate system. The physician then annotates the intended catheter entry and target tip positions, followed by documentation of the actual entry and tip positions post-procedure. This combination of automated analysis, physician input, and structured annotation creates a rich, interoperable imaging record that supports both clinical decision-making and future AI model refinement.

Figure 2: Image Adornment Workflow 

Expected Impact

By making venous anatomy effectively “visible” on routine X-rays, I-VMS supports safer insertions, earlier detection of malposition or migration, and better route selection, especially for patients who rely on durable central access for months to years. Even modest reductions in malposition could help large numbers of patients annually.

Secondary benefits include the following:

• Fewer repeat insertions
• Shorter stays
• Reduced readmissions
• Improved inter-team communication
• Stronger shared decision-making
• Clearer evidence base linking procedural choices to outcomes

Governance and Policy

 All datasets are de-identified to comply with HIPAA, but atlas-normalized coordinate data introduces a regulatory gray area. Derived datasets enable cross-patient comparisons and longitudinal analysis, yet carry re-identification risk. Policies should define permissible uses, retention and accuracy thresholds, consent requirements, and safeguards for cross-institutional sharing under HIPAA, GDPR, and emerging medical-AI rules. Because AI outputs may influence diagnosis or treatment, transparency, explainability, and clear accountability are required, along with liability frameworks for atlas-assisted decisions.

Roadmap

Planned work includes the following:

• Linking imaging hardware and encounters via FHIR Device resources
• Capturing clinician edits and overrides to create a continuous retraining loop
• Adding a graph database (e.g., Amazon Neptune) for richer querying across patients, encounters, devices, studies, and anatomical structures
• Improving projection robustness and uncertainty estimation.

Bottom Line

I-VMS unifies AI inference, open standards and clinician-in-the-loop verification to turn venous access management from fragmented documentation into a proactive, data-driven discipline. Vector-atlas projection and normalized annotations create a longitudinal, interoperable record that preserves venous options, improves safety, and enables population-level learning, while keeping raw images under local control.

See a Live Demo at the HL7 AI Challenge Winners Webinar on December 9

We invite you to attend the December 9 webinar featuring a live demo of this winner, along with Let's Talk Doc and Omni Health Nexus' The Intelligent Medical Assistant.