CBCTsynth·RT — CBCT-to-CT synthesis for radiotherapy

Turn a daily low-dose CBCT into a planning-grade synthetic CT (sCT), then compare it to a real planning CT — in a clean, modern dashboard.

Left to right: noisy CBCT input · synthetic CT (this demo's mock) · reference CT. All synthetic phantom data.

⚠️ RESEARCH / DEMO — NOT FOR DIAGNOSIS. This is a public, portfolio-style showcase. It runs a cosmetic mock on a synthetic phantom. The real model, its trained weights, the training source, and all patient data are intentionally excluded (see Privacy & Scope).

---

Why this exists

In image-guided radiotherapy, patients get a fast cone-beam CT (CBCT) on the treatment couch almost every day to check positioning. CBCT is quick and low-dose, but its image quality and HU calibration are too poor to re-plan the treatment on. A fresh planning CT means another appointment and extra dose — a real burden, especially for pediatric patients.

CBCTsynth·RT demonstrates the product around a solution to that: convert the daily CBCT into a CT-quality synthetic CT good enough for planning-grade visualization and HU analysis — no extra scan, no extra dose. This repository showcases the GUI/UX and the project goal; the underlying model is proprietary and not included here.

What the app does

1. Select a CBCT — load the bundled synthetic phantom with one click, point to a folder of DICOM files, or upload a .zip / loose files.
2. Scan info + QC — slice count, spacing, HU stats, body coverage, tissue fractions, and a deployment-QC flag (valid / borderline / invalid).
3. Generate — produce a synthetic CT and view it: axial slice slider, window/level presets (soft / bone / lung), and a synced CBCT | sCT side-by-side.
4. Compare (optional) — provide a planning CT (or the bundled demo CT). It's rigidly registered to the CBCT grid, then you get a metrics table with plain-language explanations — mean ΔHU, body MAE, structure NCC, soft-tissue ΔHU, bone P95 ΔHU, brain bone-fraction — plus a CBCT | sCT | CT side-by-side, a difference heatmap, Bland–Altman, and per-slice error.
5. Export — download the synthetic CT as a DERIVED DICOM series (geometry templated off the CBCT, tagged Research / Not for Diagnosis).

The interface uses a calm clinical palette (light blue + light orange), works in light and dark mode, and keeps heavy work cached so slice-scrubbing stays smooth.

Quickstart

# 1) install (CPU only — no GPU, no model weights needed)
pip install -r app/requirements.txt

# 2) (optional) regenerate the synthetic phantom; a copy is already committed
python sample_data/make_dummy_data.py

# 3) run the dashboard
streamlit run app/streamlit_app.py

Then click “Load bundled demo (cbct)” in step 1, Generate, and (optionally) “Load bundled demo (ct)” in step 4 to see the comparison view.

Privacy & Scope

This repo is a design / portfolio showcase, deliberately stripped of anything sensitive:

Included ✅	Excluded ❌
Dashboard GUI/UX code	Trained model weights (*.pt / *.pth)
Mock inference backend	The real model architecture & training source (dissertation IP)
Synthetic phantom generator + small committed sample	Any real patient DICOM / PHI
Metrics, visualization, DICOM export utilities	Private caches, registered/seg caches, large exports

• The "Generate" button runs a cosmetic mock (denoise + HU recalibration), clearly labeled in-app. It is not the real network and makes no clinical claim.
• All sample data is a procedurally generated phantom (ellipsoids + noise) — not anonymized real scans. See sample_data/make_dummy_data.py.
• If a private real backend (app/real_backend.py) and weights (models/*.pt) are ever present locally, the app automatically uses them instead of the mock — but neither is committed here.

Repository layout

cbctsynth-rt/
├── app/
│   ├── streamlit_app.py     # the dashboard (this is the showcase)
│   ├── pipeline.py          # public backend: DICOM load, QC, mask, export, generate router
│   ├── mock_pipeline.py     # cosmetic stand-in for the real model
│   ├── metrics.py           # HU + structure comparison metrics
│   ├── viz.py               # windowing, slice rendering, figures
│   └── requirements.txt
├── sample_data/
│   ├── make_dummy_data.py   # synthetic phantom CBCT + CT generator
│   └── demo/                # small committed phantom (cbct/ + ct/)
├── models/
│   └── MODEL_CARD.md        # describes the (absent) real models
├── docs/
│   ├── project_overview.md  # non-technical overview
│   ├── method_summary.md    # how it works, high level
│   └── metrics_explained.md # what each metric means
└── assets/screenshots/      # README visuals

Documentation

• Project overview — the problem, the impact, what this demonstrates.
• Method summary — how CBCT→sCT works, at a high level.
• Metrics explained — every number in the comparison view, in plain English.
• Model card — the real models (kept private) and their intended use.

---

CBCTsynth·RT is a research/portfolio demonstration. It is not a medical device and must not be used for diagnosis or treatment decisions.