CyberNeuro platform: Chat with Your Neuroimaging Data

Go to https://acmlab.github.io/brain-network-chart/

Examples:

1. Click on Demo button.
2. Type: Study on the correlation between Amyloid_lS_orbital_med and Tau_Global grouping by DX.
3. Click on the visualization card to enter editing mode. Type: Change dot color to red.
4. Click on download button to get SVG of the plot.
5. Type: Clustering samples into k=3 clusters given their regional Amyloid.

Requirements

The minimal requirement through above github page is Ollama.

1. Install Ollama.
2. ollama pull dcarrascosa/medgemma-1.5-4b-it:F16.
3. ollama pull gpt-oss:20b-cloud. Or pull any general model you like.

Setup on Windows (PowerShell):

[Environment]::SetEnvironmentVariable("OLLAMA_HOST", "0.0.0.0:11434", "User")
[Environment]::SetEnvironmentVariable("OLLAMA_ORIGINS", "https://acmlab.github.io,http://localhost:8000,http://127.0.0.1:8000", "User")

Then restart Ollama (quit and reopen the app).

Setup on Linux (systemd):

sudo systemctl edit ollama

Add:

[Service]
Environment="OLLAMA_HOST=0.0.0.0:11434"
Environment="OLLAMA_ORIGINS=https://acmlab.github.io,http://localhost:8000,http://127.0.0.1:8000"

Then apply:

sudo systemctl daemon-reload
sudo systemctl restart ollama
curl http://127.0.0.1:11434/api/tags

GitHub Pages runtime setup (Ollama + MCP)

When the frontend is hosted at https://acmlab.github.io/brain-network-chart/, the browser still calls local services by default:

• Ollama: http://127.0.0.1:11434 (from services/ollamaService.ts)
• Node MCP backend: http://localhost:8789 (from services/mcpService.ts)
• MCP HTTP API: http://localhost:8010 (from services/mcpService.ts)

This means each user running the GitHub Page must run local Ollama + backend + MCP server, unless you change those URLs.

If you want to chat with your brain network, run the MCP server containing available tools:

1) (Optional) Start MCP server on port 8010

cd mcp_server
uv sync
uvicorn mcp_server:http_app --host 0.0.0.0 --port 8010

2) (Optional) Start Node MCP backend on port 8789

In a new terminal:

cd cyberneuro_multi-agent-neuroimaging-analysis
FRONTEND_ORIGIN=https://acmlab.github.io PORT=8789 MCP_SERVER_URL=http://localhost:8010/mcp node backend/server.mjs

2) Open GitHub Page

• https://acmlab.github.io/brain-network-chart/
• Expected status lights in UI: Ollama green, MCP green

If either is red, verify ports 11434, 8010, 8789 and environment variables above.

Runtime Architecture

Browser UI and agents API
↓ HTTP / WebSocket
Node backend (`backend/server.mjs`)
↓ MCP
MCP Server(s)

Examples questions:

1. Click on Demo button.
2. Type: Study on the correlation between Amyloid_lS_orbital_med and Tau_Global grouping by DX.
3. Click on the visualization card to enter editing mode. Type: Change dot color to red.
4. Click on download button to get SVG of the plot.
5. Type: Clustering samples into k=3 clusters given their regional Amyloid.

Example data

See uploaded_files

Robustness test

Details refer to a2a-server/README_ROBUSTNESS_TEST.md

Biomarker queries (tabular data):

• Success rate = 93.6%
• Average time per query = 19.33s

Note on metric differences: the 93.6% figure is the end-to-end biomarker pipeline success rate (full pipeline completion + validator pass), while the 95.8% figure below is the researcher-trigger decision accuracy in the robustness test. These measure different tasks on different datasets, so the values are not expected to match.

Confusion Matrix

                 Predicted Positive    Predicted Negative
                 (Researcher Called)   (Researcher Skipped)
Actual Positive  True Positive (TP)    False Negative (FN)
(Should Call)    165                   21

Actual Negative  False Positive (FP)   True Negative (TN)
(Shouldn't Call) 0                     314

Metrics:

• Precision = TP/(TP+FP) = 165/(165+0) = 100.0%
• Recall = TP/(TP+FN) = 165/(165+21) = 88.7%
• Accuracy = (TP+TN)/(Total) = (165+314)/500 = 95.8%

Tool calling times

✅ connectivity analysis          - 68 queries (13.6%)
✅ network hub identification     - 61 queries (12.2%)
✅ cross-frequency coupling       - 56 queries (11.2%)
✅ functional connectivity        - 56 queries (11.2%)
✅ normative analysis             - 51 queries (10.2%)
✅ wavelet decomposition          - 47 queries (9.4%)
✅ CFC wavelet analysis           - 46 queries (9.2%)
✅ system segregation analysis    - 42 queries (8.4%)
✅ hub detection                  - 38 queries (7.6%)
✅ growth curve modeling          - 35 queries (7.0%)