Skip to content

Imaging Intelligence Agent -- Demo Guide

UI-driven walkthrough for demonstrating the Imaging Intelligence Agent on DGX Spark.

All live demo interaction uses the Streamlit Imaging Workbench -- no terminal commands during the presentation.

License: Apache 2.0 | Date: March 2026

Demo Overview

Parameter Value
Total Duration 25 minutes
Hardware NVIDIA DGX Spark (GB10, 128 GB unified)
Total Vectors 876 seed vectors across 10 owned collections
Streamlit URL http://localhost:8525
API URL http://localhost:8524
LLM Claude Sonnet 4.6 (Anthropic)
Milvus http://localhost:19530
Collections 10 owned + 1 shared genomic_evidence
Workflows 6 (CT Head Hemorrhage, CT Chest Lung Nodule, CT Coronary Angiography, CXR Rapid Findings, MRI Brain MS Lesion, MRI Prostate PI-RADS)

What the Audience Will See

  1. Four pre-loaded clinical demo cases run through imaging AI workflows with mock-realistic results
  2. Cross-modal genomic enrichment linking imaging findings to genetic risk factors via the shared genomic_evidence collection
  3. Image Gallery with AI-annotated CXR showcase, cross-modality gallery, and 3D volume slice viewer
  4. Interactive Plotly network graph in Patient 360 showing gene-finding relationships
  5. Multi-collection RAG queries across 10 imaging-specific knowledge bases with Claude-synthesized answers and citations, with pre-filled example queries in Evidence Explorer and Protocol Advisor
  6. Automatic comparative analysis triggered by natural language -- "CT vs MRI" produces structured comparison tables
  7. Patient-specific imaging protocol recommendations with AI-optimized dose reduction
  8. A landscape of 50 FDA-cleared and research AI devices filterable by modality and clinical task
  9. Dose intelligence showing 36% average radiation dose reduction across 20 protocols via DLIR
  10. 6-step pipeline animation in Workflow Runner with annotated AI images
  11. Sidebar guided tour with 9-step demo flow
  12. Professional reports exported as Markdown, JSON, NVIDIA-themed PDF, and FHIR R4 DiagnosticReport Bundles

Pre-Demo Setup

Step 1: Verify Milvus Running

curl -s http://localhost:19530/v1/vector/collections
# Expected: JSON response listing imaging collections

Step 2: Verify Docker Containers

docker ps | grep imaging
# Expected: imaging-intelligence-agent containers running (API + Streamlit)

Step 3: Health Check

curl -s http://localhost:8524/health | python3 -m json.tool

Expected response:

{
  "status": "healthy",
  "collections": {
    "imaging_literature": 50,
    "imaging_trials": 40,
    "imaging_findings": 50,
    "imaging_protocols": 40,
    "imaging_devices": 50,
    "imaging_anatomy": 30,
    "imaging_benchmarks": 40,
    "imaging_guidelines": 40,
    "imaging_report_templates": 50,
    "imaging_datasets": 50
  },
  "total_vectors": 440,
  "nim_services": {
    "vista3d": "mock",
    "maisi": "mock",
    "vila_m3": "mock",
    "llm": "anthropic"
  }
}

All 10 collections should have non-zero counts. NIM services will show mock unless NVIDIA API keys are configured; the LLM should show anthropic.

Step 4: Verify Streamlit

curl -s http://localhost:8525/_stcore/health
# Expected: "ok"

Step 5: Verify ANTHROPIC_API_KEY

echo $ANTHROPIC_API_KEY | head -c 10
# Expected: sk-ant-api (first 10 characters of your key)

Step 6: Open Browser Tabs

Before starting the live demo, open the following tabs in your browser so you can switch between them without typing URLs:

Tab URL Used In
Imaging Agent UI http://localhost:8525 All steps
Imaging Agent API docs http://localhost:8524/docs Reference only
Landing Page http://localhost:8080 Opening context

Route A: Guided Demo Cases (15 minutes)

These four cases demonstrate the full imaging AI workflow: detection, classification, measurement, and cross-modal genomic enrichment.

All cases use pre-loaded mock data that runs without live DICOM images or NIM GPU services.

Step 1: Emergency Stroke (5 minutes)

Click: the Workflow Runner tab (the second tab) in the Streamlit UI at http://localhost:8525.

Select: "DEMO-001: Emergency Stroke: Acute Intracranial Hemorrhage" from the demo case dropdown.

Expected result: The clinical scenario populates:

62-year-old male presents to ED with sudden onset severe headache, left-sided weakness, and slurred speech. GCS 12. History of uncontrolled hypertension and type 2 diabetes. BP 195/110 on arrival. Non-contrast CT head ordered stat for acute stroke workup.

Click: the Run button.

Expected result: The workflow executes the ct_head_hemorrhage pipeline and displays:

  • Hemorrhage Detection: Intraparenchymal hemorrhage detected
  • Location: Right basal ganglia extending to internal capsule
  • Volume: 28.5 mL (ABC/2 method)
  • Midline Shift: 4.8 mm
  • Max Thickness: 42.0 mm
  • Hounsfield Units: Mean 65, Max 82
  • Surrounding Edema: 12.1 mL
  • Intraventricular Extension: Present
  • Fisher Grade: 3
  • Severity Classification: Critical

Click: the Patient 360 tab (the eighth tab).

Show: Cross-modal genomic enrichment results linking the hemorrhage finding to genetic risk factors:

  • APOE -- e4 allele associated with lobar hemorrhage risk and worse outcomes
  • COL3A1 -- Variants linked to vascular fragility
  • ACE -- Insertion/deletion polymorphism affects cerebrovascular risk

Talking points:

  • "AI detected the hemorrhage in under 90 seconds -- faster than radiologist page response time."
  • "Automatic midline shift measurement eliminates subjective assessment. 4.8 mm shift is clinically significant."
  • "Cross-modal genomic enrichment identifies genetic risk factors for hemorrhage recurrence -- APOE, COL3A1, and ACE polymorphisms."
  • "Structured report with ICH Score automatically generated for neurosurgery consult."

Step 2: Lung Cancer Screening (5 minutes)

Click: the Workflow Runner tab.

Select: "DEMO-002: Lung Cancer Screening: Suspicious Nodule Detection" from the demo case dropdown.

Expected result: The clinical scenario populates:

58-year-old female, 30 pack-year smoking history, undergoing annual low-dose CT lung cancer screening per USPSTF guidelines. Prior screening CT 12 months ago showed a 6mm ground-glass nodule in the right upper lobe (Lung-RADS 3). Follow-up scan ordered.

Click: the Run button.

Expected result: The workflow executes the ct_chest_lung_nodule pipeline and displays two nodules:

  • Nodule 1 (Primary):

    • Location: Right upper lobe, posterior segment
    • Type: Part-solid, spiculated margin
    • Size: 18 mm (long axis) x 14 mm (short axis)
    • Solid component: 10 mm
    • Volume: 1,890 mm3 (prior: 680 mm3)
    • Volume doubling time: 245 days
    • Density: -450 HU
    • Lung-RADS: 4B -- Tissue sampling recommended

  • Nodule 2 (Incidental):

    • Location: Left lower lobe, superior segment
    • Type: Ground-glass, smooth margin
    • Size: 5 mm
    • Volume: 52 mm3
    • Density: -650 HU
    • Lung-RADS: 2 -- Benign appearance, no intervention

Show: Genomic enrichment panel linking to driver mutations:

  • EGFR -- Mutations found in 15-50% of lung adenocarcinomas
  • ALK -- Rearrangements in 3-7% of NSCLC
  • ROS1, KRAS, BRAF, MET -- Additional targetable driver mutations

Talking points:

  • "AI tracked nodule growth from 6 mm to 18 mm over 12 months -- volume doubling time of 245 days is concerning for malignancy."
  • "Automatic Lung-RADS 4B classification triggers tissue sampling recommendation per ACR guidelines."
  • "The second nodule (5 mm GGN, Lung-RADS 2) is appropriately classified as benign -- no intervention needed."
  • "Cross-modal genomic query identifies targetable driver mutations: EGFR, ALK, ROS1, KRAS."
  • "Complete screening-to-genomics pipeline: low-dose CT -> AI detection -> Lung-RADS classification -> molecular profiling -> treatment planning."

Step 3: Cardiac Workup (5 minutes)

Click: the Workflow Runner tab.

Select: "DEMO-003: Cardiac Workup: Coronary Artery Disease Assessment" from the demo case dropdown.

Expected result: The clinical scenario populates:

55-year-old male presents with exertional chest pain and dyspnea on exertion for 3 weeks. Family history of premature CAD (father MI at age 50). BMI 28, total cholesterol 265 mg/dL, LDL 185 mg/dL. Stress test equivocal. Coronary CT angiography ordered for definitive evaluation.

Click: the Run button.

Expected result: The workflow executes the ct_coronary_angiography pipeline and displays:

  • Calcium Score: 385 Agatston (92nd percentile for age/sex)
  • CAD-RADS Classification: 4A
  • Vessel Assessment:
    • LAD (proximal): 72% stenosis, mixed plaque -- significant
    • LAD (mid): 40% stenosis, calcified plaque
    • LCx (proximal): 30% stenosis, calcified plaque
    • RCA (proximal): 15% stenosis, no plaque
    • Left Main: 0% stenosis, no plaque
  • High-Risk Plaque Features: Low-attenuation plaque, positive remodeling
  • Ejection Fraction Estimate: 55%
  • Severity Classification: Urgent

Show: Genomic enrichment panel linking to familial hypercholesterolemia genes:

  • LDLR -- Variants cause 60-80% of familial hypercholesterolemia cases
  • PCSK9 -- Gain-of-function variants increase LDL cholesterol
  • APOB -- Defective ligand binding causes FH
  • LPA -- Elevated Lp(a) is an independent cardiovascular risk factor
  • 9p21 -- Strongest common CAD risk variant locus

Talking points:

  • "Calcium score of 385 (92nd percentile for age and sex) -- high atherosclerotic burden quantified in seconds."
  • "AI detected 72% LAD stenosis with high-risk plaque features: low-attenuation plaque and positive remodeling."
  • "CAD-RADS 4A classification automatically generated with guideline-concordant follow-up recommendations."
  • "Cross-modal genomic enrichment identifies familial hypercholesterolemia genes -- LDLR, PCSK9, APOB -- family cascade screening is indicated."
  • "Complete cardiac loop: CTA -> calcium score -> stenosis grading -> genomics -> risk management."

Route B: Interactive Exploration (10 minutes)

This route demonstrates the breadth of the knowledge base, protocol intelligence, dose optimization, and export capabilities.

Step 4: Evidence Explorer (3 minutes)

Click: the Evidence Explorer tab (the first tab).

Type this query in the "Ask a question" text input:

What AI models are used for intracranial hemorrhage detection?

Click: the Ask button.

Expected result:

  • A markdown answer appears identifying specific models and approaches: CNN-based classifiers, U-Net segmentation, attention-based architectures, and FDA-cleared commercial solutions.
  • Evidence sources display badges from multiple collections: imaging_literature, imaging_devices, imaging_benchmarks, imaging_findings.
  • Each source shows its cosine similarity score and a text excerpt.
  • Processing time caption appears at the bottom.

Show: The evidence count -- 20+ sources retrieved across multiple collections in a single query.

Type this comparative query in the text input:

CT vs MRI for stroke detection

Click: the Ask button.

Expected result: Claude detects the comparative keywords ("vs") and produces a structured comparison table:

Dimension CT MRI
Speed Minutes 30-60 minutes
Sensitivity (acute hemorrhage) >95% Lower for acute blood
Sensitivity (ischemic stroke) Limited (<6h) High (DWI within minutes)
Availability Ubiquitous, 24/7 Limited off-hours
Cost Lower Higher
Guideline Role First-line for acute stroke Follow-up and characterization

Talking points:

  • "One question searched all 10 collections in parallel -- literature, devices, benchmarks, and findings all contributed evidence."
  • "The comparative query was auto-detected. Claude parsed 'CT' and 'MRI' as two entities and ran dual retrieval."
  • "Every claim is backed by a citation with cosine similarity scores."

Step 5: Protocol Advisor (2 minutes)

Click: the Protocol Advisor tab (the third tab).

Type this value in the clinical indication input:

acute chest pain, rule out pulmonary embolism

Enter: Patient age = 45, Patient weight = 80 kg.

Click: the Recommend Protocol button.

Expected result: A protocol recommendation card displays:

  • Protocol: CT Pulmonary Angiography (CTPA)
  • Modality: CT
  • Contrast Agent: Iodinated IV contrast (100 mL, bolus tracking)
  • Estimated Dose: Optimized CTDIvol with AI dose reduction percentage
  • Duration: Estimated scan time
  • AI Optimization Notes: DLIR (Deep Learning Image Reconstruction) recommendations, kVp optimization for patient weight
  • Alternatives: V/Q scan (if contrast allergy), D-dimer follow-up (if low pretest probability)

Talking points:

  • "Patient-specific dose adjustment based on age and weight -- the protocol adapts to the individual."
  • "AI optimization notes recommend DLIR for noise reduction at lower dose levels."
  • "Alternatives are provided for patients with contrast allergies -- following the ALARA principle."

Step 6: Dose Intelligence (2 minutes)

Click: the Dose Intelligence tab (the fifth tab).

Expected result: The dose comparison dashboard displays:

  • Summary Statistics: 20 protocols analyzed, average dose reduction, max/min reduction by modality.
  • Standard vs AI-Optimized Dose Chart: Side-by-side bar chart comparing CTDIvol (mGy) for standard protocols vs AI-optimized protocols.
  • Protocol Table: All 20 protocols with standard dose, AI-optimized dose, reduction percentage, technique used, and image quality assessment.

Show: The average dose reduction metric.

Expected result: ~36% average dose reduction across all 20 protocols. Individual protocols range from 15-55% reduction depending on body region and technique.

Talking points:

  • "DLIR (Deep Learning Image Reconstruction) reduces radiation dose by an average of 36% without sacrificing diagnostic image quality."
  • "The highest reductions are in pediatric protocols and repeat imaging studies."
  • "Image quality is maintained or improved -- AI reconstruction compensates for lower photon counts."
  • "This directly supports the ALARA principle: As Low As Reasonably Achievable."

Step 7: Device & AI Ecosystem (1 minute)

Click: the Device & AI Ecosystem tab (the fourth tab).

Expected result: A searchable, filterable catalog of AI devices displays with columns for device name, manufacturer, modality, clinical task, regulatory status, and performance metrics.

Select: Modality filter = CT, Task filter = detection.

Expected result: The table filters to show CT-specific AI detection devices -- triage tools, hemorrhage detectors, pulmonary embolism detectors, nodule detection, and fracture detection solutions.

Show: The total device count: 50 FDA-cleared and research AI devices across all modalities.

Talking points:

  • "A landscape of 50+ AI devices spanning CT, MRI, X-ray, ultrasound, mammography, and PET."
  • "Each device is catalogued with regulatory status, clinical task, and published performance benchmarks."
  • "This collection enables evidence-based evaluation of AI tools for clinical deployment."

Step 8: Reports & Export (2 minutes)

Click: the Reports & Export tab (the sixth tab).

Click: the Generate Report button to create a report from the last query.

Expected result: A formatted clinical report renders inline with sections:

  • Clinical Question -- the original query
  • Analysis -- Claude-synthesized answer with structured findings
  • Evidence Citations -- numbered list with collection badges, document IDs, and relevance scores

Show: The four export format buttons:

  1. Markdown -- Copy-paste ready for clinical notes
  2. JSON -- Structured data for programmatic consumption
  3. PDF -- NVIDIA-themed report with green headers and professional formatting
  4. FHIR R4 -- DiagnosticReport Bundle with SNOMED CT and LOINC coding

Click: the PDF button.

Expected result: A PDF downloads with NVIDIA-branded formatting, clinical question, analysis, evidence citations with relevance scores, and a disclaimer footer.

Click: the FHIR R4 button.

Expected result: A FHIR R4 DiagnosticReport Bundle renders as JSON, including resourceType: "Bundle", DiagnosticReport resource with LOINC-coded sections, and Observation resources for each finding.

Talking points:

  • "Four export formats for different integration needs: Markdown for sharing, JSON for APIs, PDF for clinical documentation, FHIR R4 for EHR interoperability."
  • "The FHIR R4 bundle uses SNOMED CT and LOINC coding for standards-based interoperability."
  • "PDF reports include all evidence citations with relevance scores -- fully auditable AI-assisted findings."

Troubleshooting

No Collections Found

If the health check shows zero vectors or collections are missing:

cd ai_agent_adds/imaging_intelligence_agent/agent
python3 scripts/setup_collections.py --drop-existing
python3 scripts/seed_literature.py
python3 scripts/seed_trials.py
python3 scripts/seed_findings.py
python3 scripts/seed_protocols.py
python3 scripts/seed_devices.py
python3 scripts/seed_anatomy.py
python3 scripts/seed_benchmarks.py
python3 scripts/seed_guidelines.py
python3 scripts/seed_report_templates.py
python3 scripts/seed_datasets.py

RAG Returns Empty Answers

If queries return evidence but the answer is empty or generic:

# Verify ANTHROPIC_API_KEY is set and valid
echo $ANTHROPIC_API_KEY | head -c 10
# Expected: sk-ant-api

# Test Claude directly
curl -s https://api.anthropic.com/v1/messages \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model": "claude-sonnet-4-6", "max_tokens": 10, "messages": [{"role": "user", "content": "Hi"}]}'

Workflow Failed

If a demo case workflow fails to execute:

# Check Docker containers are healthy
docker ps | grep imaging
docker compose logs imaging-api

# Verify workflow registry
curl -s http://localhost:8524/demo-cases | python3 -m json.tool
# Expected: 4 demo cases listed

Port In Use

If services fail to start due to port conflicts:

# Check what is using the API port
ss -tlnp | grep 8524

# Check what is using the Streamlit port
ss -tlnp | grep 8525

# Check running Docker containers
docker ps --format "table {{.Names}}\t{{.Ports}}"

Streamlit UI Not Loading

# Check Streamlit container status
docker compose ps

# View Streamlit logs
docker compose logs streamlit-imaging

# Restart Streamlit service
docker compose restart streamlit-imaging

PDF Export Fails

PDF generation requires ReportLab. If PDF export returns a 501 error:

pip install reportlab

Quick Reference

URLs

Resource URL
Imaging Agent UI (Streamlit) http://localhost:8525
Imaging Agent API docs (Swagger) http://localhost:8524/docs
Landing Page http://localhost:8080
Milvus http://localhost:19530
Attu (Milvus UI) http://localhost:8000
Grafana Monitoring http://localhost:3000

Collections Reference

Collection Purpose Approx. Vectors
imaging_literature Published research papers and reviews 50
imaging_trials ClinicalTrials.gov AI-in-imaging records 40
imaging_findings Imaging finding templates and patterns 50
imaging_protocols Acquisition protocols and parameters 40
imaging_devices FDA-cleared AI/ML medical devices 50
imaging_anatomy Anatomical structure references 30
imaging_benchmarks Model performance benchmarks 40
imaging_guidelines Clinical practice guidelines (ACR, RSNA) 40
imaging_report_templates Structured radiology report templates 50
imaging_datasets Public imaging datasets (TCIA, PhysioNet) 50
genomic_evidence Shared genomic variants (read-only, Stage 1) 3,561,170

Key curl Examples

# Health check
curl -s http://localhost:8524/health | python3 -m json.tool

# List collections
curl -s http://localhost:8524/collections | python3 -m json.tool

# List demo cases
curl -s http://localhost:8524/demo-cases | python3 -m json.tool

# Run a demo case
curl -s -X POST http://localhost:8524/demo-cases/DEMO-001/run | python3 -m json.tool

# Protocol recommendation
curl -s -X POST http://localhost:8524/protocol/recommend \
  -H "Content-Type: application/json" \
  -d '{
    "indication": "acute chest pain, rule out pulmonary embolism",
    "patient_age": 45,
    "patient_weight_kg": 80
  }' | python3 -m json.tool

# Dose summary
curl -s http://localhost:8524/dose/summary | python3 -m json.tool

Appendix: API Reference

All endpoints below are served by the FastAPI server at http://localhost:8524. These are provided for programmatic integration, automated testing, and scripting -- not for live demo use. Use the Streamlit UI at port 8525 for demos.

Interactive API documentation is available at http://localhost:8524/docs (Swagger UI).

Health and Status

# Health check -- all collections, vectors, and NIM service status
curl -s http://localhost:8524/health | python3 -m json.tool

# List all collections with vector counts and labels
curl -s http://localhost:8524/collections | python3 -m json.tool

# Knowledge graph statistics
curl -s http://localhost:8524/knowledge/stats | python3 -m json.tool

# Prometheus metrics
curl -s http://localhost:8524/metrics

Demo Cases

# List all demo cases
curl -s http://localhost:8524/demo-cases | python3 -m json.tool

# Run a specific demo case (DEMO-001, DEMO-002, DEMO-003, or DEMO-004)
curl -s -X POST http://localhost:8524/demo-cases/DEMO-001/run | python3 -m json.tool

# Response includes: workflow_result, genomic_context, talking_points

RAG Queries

# Full RAG query (multi-collection retrieval + LLM synthesis)
curl -s -X POST http://localhost:8524/query \
  -H "Content-Type: application/json" \
  -d '{
    "question": "What AI models are used for intracranial hemorrhage detection?",
    "modality": "ct",
    "body_region": "head",
    "top_k": 5,
    "include_genomic": true
  }' | python3 -m json.tool

# Evidence-only search (retrieval only, no LLM synthesis)
curl -s -X POST http://localhost:8524/search \
  -H "Content-Type: application/json" \
  -d '{
    "question": "lung nodule detection deep learning",
    "modality": "ct",
    "top_k": 10
  }' | python3 -m json.tool

# Cross-collection entity linking
curl -s -X POST http://localhost:8524/find-related \
  -H "Content-Type: application/json" \
  -d '{
    "entity": "pulmonary embolism",
    "top_k": 5
  }' | python3 -m json.tool

Protocol Optimization

# Recommend imaging protocol
curl -s -X POST http://localhost:8524/protocol/recommend \
  -H "Content-Type: application/json" \
  -d '{
    "indication": "acute chest pain, rule out pulmonary embolism",
    "patient_age": 45,
    "patient_weight_kg": 80,
    "patient_sex": "male",
    "contrast_allergy": false,
    "pregnancy": false
  }' | python3 -m json.tool

Dose Intelligence

# Get all dose comparison data (20 protocols)
curl -s http://localhost:8524/dose/reference | python3 -m json.tool

# Get dose comparison for a specific protocol
curl -s http://localhost:8524/dose/comparison/chest | python3 -m json.tool

# Get summary statistics (avg/max/min reduction, by modality)
curl -s http://localhost:8524/dose/summary | python3 -m json.tool

Report Export

# Markdown report
curl -s -X POST http://localhost:8524/reports/generate \
  -H "Content-Type: application/json" \
  -d '{
    "question": "What is the optimal imaging protocol for suspected stroke?",
    "modality": "ct",
    "format": "markdown"
  }' | python3 -m json.tool

# JSON report
curl -s -X POST http://localhost:8524/reports/generate \
  -H "Content-Type: application/json" \
  -d '{
    "question": "What is the optimal imaging protocol for suspected stroke?",
    "format": "json"
  }' | python3 -m json.tool

# PDF report (binary download)
curl -s -X POST http://localhost:8524/reports/generate \
  -H "Content-Type: application/json" \
  -d '{
    "question": "What is the optimal imaging protocol for suspected stroke?",
    "format": "pdf"
  }' --output imaging_report.pdf

HCLS AI Factory -- Apache 2.0 | March 2026