Evaluate imaging AI inside your own PACS workflow

PACS-AI is a research platform from the HeartWise AI Lab at the Montreal Heart Institute. It pulls studies from your PACS, runs AI models in containers on servers you control, and shows the results in a DICOM viewer where clinicians review and rate them.

The demo needs a free account. Create an account

Research use only. PACS-AI is not approved as a medical device, and the models on this site are research models that are not cleared for clinical use. Running them on patient studies requires a research protocol approved by your institution.

Silent screen recording of the PACS-AI viewer, about 19 seconds long. CathEF estimates left ventricular ejection fraction (LVEF) from coronary angiograms. It is a research model, and its estimate is not a clinical measurement.

How a study moves through PACS-AI

Everything below runs on servers your institution controls. Each step shows what you will find in the system and links to the documentation page that covers it.

  1. Connect your PACS

    PACS-AI queries your PACS over DICOM and copies matching studies into its own Orthanc server. By default, copies expire from that cache after 24 hours.

    In the system

    • C-FIND
    • C-MOVE
    • Orthanc
    • ORTHANC_LOCAL_CACHE_EXPIRATION_IN_HOURS=24

    Who: Hospital IT

    PACS integration
  2. Package a model

    Each model runs in its own Docker container. The container declares the studies it accepts, publishes a Model Facts label, and answers prediction requests.

    In the system

    • model_info.json
    • model_facts.json
    • POST /inference/predict
  3. Run models on new studies

    Ingestion jobs look for studies that match a modality and a schedule, wait until each study has finished arriving, then run the assigned models. A clinician can also run a model on one study from the viewer.

    In the system

    1. DISCOVERED
    2. STABLE
    3. RETRIEVED
    4. PROCESSING
    5. PROCESSED

    Who: Research team or administrator

    Automated inference pipeline
  4. Review results next to the images

    Results open in an OHIF-based viewer as image annotations, an HTML or PDF report, JSON, or the model's own web view. A result can also be saved back to the study in your PACS as a new series.

    In the system

    • OHIF_ANNOTATIONS
    • HTML
    • PDF
    • JSON
    • WEB_APP

    Who: Clinicians

    Output modes
  5. Collect clinician feedback

    Readers approve or reject each result and answer that model's short questionnaire. Sign-ins, account changes and results saved to the PACS are recorded in an audit log.

    In the system

    • APPROVE
    • REJECT
    • Elasticsearch audit log

    Who: Clinicians and research team

    Security and audit logging

Where the data goes

Images and model results are processed on the PACS-AI server inside your network. A standard install also relies on a few outside services. They are listed here so your privacy and security teams can review them.

Inside your institution's network

Your PACS

Source of the studies

DICOM

PACS-AI server

  • Orthanc cache for incoming studies
  • Control plane and job database
  • Model containers on an NVIDIA GPU host
  • Audit log (Elasticsearch)

HTTPS

Clinician's browser

Viewer served by your PACS-AI server

Outside services used by a standard install

  • Google Firebase

    Sign-in, user accounts and organization settings (Firebase Authentication and Firestore). Firebase Authentication runs in US data centres.

  • Mailgun

    Account email: verification, invitations and password resets.

  • Docker Hub, Hugging Face

    Platform images and model weights, downloaded at install or build time.

  • Cloudflare Turnstile

    Bot protection for sign-up and sign-in.

Diagram: the PACS, the PACS-AI server and the clinician's browser are inside the institution's network. Firebase, Mailgun, Docker Hub, Hugging Face and Cloudflare Turnstile are outside services.
  • PACS-AI does not de-identify studies. Images keep their identifiers in the cache, so de-identify upstream if your protocol requires it.
  • TLS protects browser and API traffic. Encryption at rest relies on your server's disk encryption. DICOM-TLS between the PACS and PACS-AI is supported but off by default.
  • Fully offline installs are only partly supported: sign-in still needs to reach Firebase.
  • There is no SOC 2, HITRUST or ISO 27001 attestation today. A compliance program covering HIPAA, GDPR and Quebec's Law 25 is in development with external counsel.

Research behind the models

Explore studies and preprints behind models developed by or available through PACS-AI.

Each paper reports on a model, not on PACS-AI itself. Results can differ at your site, so evaluate a model locally before relying on it.

  • JAMA Cardiology, 2023

    CathEF Study: AI-Powered Cardiac Function Analysis

    Robert Avram et al.

    A video-based deep neural network that estimates left ventricular ejection fraction from standard left coronary angiograms, compared against transthoracic echocardiography and externally validated at the University of Ottawa Heart Institute.

    Read the paper
  • European Heart Journal – Digital Health, 2026

    DeepRV: Automated Right Ventricular Function from Coronary Angiograms

    Fatima Zahra Fawzi et al.

    A video-based deep learning model that assesses right ventricular systolic function directly from routine coronary angiograms, removing the need for prior imaging. Developed at the Montreal Heart Institute, externally validated at UCSF, and deployed prospectively during primary PCI for STEMI.

    Read the paper
  • arXiv preprint, 2026

    DeepCORO-CLIP: A Multi-View Foundation Model for Coronary Angiography

    Sarra Harrabi et al.

    A multi-view foundation model trained with video-text contrastive learning for comprehensive coronary angiography interpretation, with external validation and publicly released code and weights.

    Read the paper
  • Nature, 2026

    EchoPrime: A Multi-View Vision-Language Model for Echocardiography

    Milos Vukadinovic et al.

    A view-informed vision-language foundation model for echocardiography that interprets every video in a complete study rather than a single view, using contrastive learning and retrieval-augmented interpretation.

    Read the paper
  • JAMA, 2025

    PanEcho: Complete AI-Enabled Echocardiography Interpretation

    Gregory Holste et al.

    A view-agnostic multi-task model that performs the full range of echocardiographic reporting tasks from a single pass, validated internally and across four external cohorts, and publicly released.

    Read the paper
  • arXiv preprint, 2024

    CardioSyntax: End-to-End SYNTAX Score Prediction from Angiography

    Alexander Ponomarchuk et al.

    An automatic method for estimating the SYNTAX score of coronary disease severity from multi-view angiography video, published together with a purpose-built dataset and benchmark.

    Read the paper
View all publications

Research access

Free for research and evaluation

Available at no cost to clinicians, researchers, and institutions evaluating AI in medical imaging.

Research Use

Free

What you get

  • Source code for the viewer (MIT licence) and the control plane (GNU GPL v3) on GitHub
  • A public demo with a curated anonymized dataset, hosted in Quebec (free account required)
  • Documentation for installing, operating and securing your own instance
  • A model template and example model containers to adapt

Models are not approved by the FDA or Health Canada. For research purposes only.

What your institution provides

  • A research protocol approved by your research ethics board
  • IT staff to connect your PACS and run a Linux server with an NVIDIA GPU
  • Clinical leads accountable for how each model and its results are used
  • Google Firebase and Mailgun accounts for sign-in and account email

Good to know

  • The execution service that runs model containers is in a private repository. Access is granted per project.
  • Support is best effort, Monday to Friday on Eastern Time, through GitHub issues or a Slack channel set up for your project.

Frequently asked questions

Can't find your answer? Contact the team

PACS-AI connects to your PACS over DICOM, copies matching studies into a local cache, and runs AI models packaged as containers on your servers. Results open in an OHIF-based viewer, where clinicians review and rate them. The section How a study moves through PACS-AI, above, covers each step.

With support from

CIFAR IVADO

Get updates on PACS-AI

Occasional news about releases, models and publications.