AXIO PREDICT

Cross-platform desktop application for lung cancer risk prediction from low-dose CT scans, powered by the Sybil v1.6.0 deep learning model.

AXIO PREDICT provides a local, privacy-preserving graphical interface for clinicians and researchers to load DICOM or PNG CT series, run the validated Sybil ensemble model, visualise attention maps, and audit prediction quality — all without sending patient data to any external server.

---

Screenshots

Landing screen	Dark mode

Risk Assessment Report	Audit Toolkit

---

Features

• 6-year risk scores — calibrated probabilities of lung cancer development at 1–6 years from scan date
• DICOM & PNG support — axial LDCT series, auto-sorted by z-position; .dcm, .dicom, .ima
• Attention map visualisation — GIF overlays highlighting contributing slices and pixels
• Audit Toolkit — proxy attention diagnostics: outside-body attention, boundary dominance, peripheral-ring attention, superior/inferior asymmetry, left-right asymmetry, top-1/top-3 slice dominance
• Local checkpoints — model weights never leave your machine
• Cross-platform — macOS (Apple Silicon + Intel), Windows, Linux
• Offline inference — no internet required after initial model download

---

Risk Classification

Score	Meaning
Year 1–6	Calibrated probability of lung cancer within N years

Level	Threshold
Low	< 5%
Moderate	5–15%
High	> 15%

---

Prerequisites

Tool	Version	Purpose
Python	3.8–3.12	Sybil inference backend
Node.js	≥ 16	Electron runtime
npm	≥ 8	Package management

---

Installation

1. Clone the repository

git clone https://github.com/godsonj64/AXIOPREDICT.git
cd AXIOPREDICT

2. Set up the Python environment

macOS / Linux:

bash setup.sh

Windows:

setup.bat

This creates a .venv/ virtual environment and installs Flask + Sybil from the bundled source.

3. Install Electron dependencies

cd electron_app
npm install

4. Download model checkpoints

Download the Sybil ensemble weights from one of:

• GitHub Releases: https://github.com/reginabarzilaygroup/Sybil/releases
• Google Drive: https://drive.google.com/drive/folders/1nBp05VV9mf5CfEO6W5RY4ZpcpxmPDEeR

Place all .ckpt files and the calibrator .json in a single directory, e.g.:

~/sybil_checkpoints/
├── 28a7cd44f5bcd3e6cc760b65c7e0d54d.ckpt
├── 56ce1a7d241dc342982f5466c4a9d7ef.ckpt
├── 624407ef8e3a2a009f9fa51f9846fe9a.ckpt
├── 64a91b25f84141d32852e75a3aec7305.ckpt
├── 65fd1f04cb4c5847d86a9ed8ba31ac1a.ckpt
└── sybil_ensemble_simple_calibrator.json

5. Run

npm start

In the app: choose Local Checkpoints, browse to your checkpoints directory, click Load Model.

---

Building a Distributable

cd electron_app

# macOS (universal DMG)
npm run dist:mac

# Windows (NSIS installer)
npm run dist:win

# Linux (AppImage)
npm run dist:linux

The Python backend and Sybil source are bundled as extraResources in the packaged app.

---

Project Structure

AXIOPREDICT/
├── setup.sh / setup.bat      ← environment setup
├── python_backend/
│   ├── server.py             ← Flask HTTP API wrapping Sybil
│   └── requirements.txt
├── sybil-source/             ← Sybil v1.6.0 source (unmodified)
│   └── sybil/
│       ├── model.py
│       ├── serie.py
│       ├── models/           ← SybilNet architecture
│       ├── loaders/          ← DICOM/PNG loading pipeline
│       └── utils/            ← visualisation, metrics
└── electron_app/
    ├── package.json
    └── src/
        ├── main.js           ← Electron main process, spawns Python backend
        ├── preload.js        ← secure IPC bridge
        └── index.html        ← full UI (HTML/CSS/JS, no framework)

---

Audit Toolkit

AXIO PREDICT v2.0.1 adds a paper-inspired audit toolkit informed by the 2026 preprint Auditing Sybil: Explaining Deep Lung Cancer Risk Prediction Through Generative Interventional Attributions.

Proxy diagnostics included:

• Outside-body attention leakage
• Boundary-dominant attention
• Peripheral-ring attention
• Superior vs. inferior slice attention imbalance
• Left-right asymmetry
• Top-1 / top-3 slice dominance

Scope: These are non-causal proxy checks computed from existing attention maps. They are review aids, not proof of lesion-level causality. The underlying Sybil prediction path, model weights, calibration, and checkpoint handling are unchanged.

---

Citation

If you use AXIO PREDICT in your work, please cite both this application and the underlying Sybil model.

AXIO PREDICT

@software{johnson2026axiopredict,
  author    = {Johnson, Godson},
  title     = {{AXIO PREDICT}: A Desktop Application for Lung Cancer Risk
               Prediction from Low-Dose {CT} Scans},
  year      = {2026},
  version   = {2.0.1},
  url       = {https://github.com/godsonj64/AXIOPREDICT},
  license   = {MIT}
}

Sybil (underlying model)

@article{mikhael2023sybil,
  author    = {Mikhael, Peter G and Wohlwend, Jeremy and Yala, Adam and
               Karstens, Ludvig and Xiang, Justin and Takigami, Angelo K and
               Bourgouin, Patrick P and Chan, PuiYee and Mrah, Sofiane and
               Amayri, Wael and Juan, Yu-Hsiang and Yang, Cheng-Ta and
               Wan, Yung-Liang and Balakrishnan, Guha and
               Sequist, Lecia V and Fintelmann, Florian J and Barzilay, Regina},
  title     = {Sybil: A Validated Deep Learning Model to Predict Future Lung
               Cancer Risk from a Single Low-Dose Chest Computed Tomography},
  journal   = {Journal of Clinical Oncology},
  year      = {2023},
  doi       = {10.1200/JCO.22.01345},
  publisher = {Wolters Kluwer Health}
}

---

License

MIT © 2026 Godson Johnson

---

Disclaimer

FOR RESEARCH USE ONLY. Not validated for clinical diagnosis or treatment decisions. Always involve qualified medical professionals in clinical workflows.