Evidence, not images

The trust layer for CT restoration AI.

One engine — physics-grounded degradation ⇄ do-no-harm restoration ⇄ task-based evidence — powering three products: prove it, use it, teach it.

● 1 validated core RSNA · Mayo · TotalSegmentator Phase 1: research / QA / education only — not diagnosis
The shared core

One engine, already validated in the notebook

The same two functions — degrade() and restore() — plus one scorer feed all three surfaces. A scan degraded to teach a student is generated by the exact pipeline that stress-tests a model and trains the restorer.

Degradephysics-grounded synthetic noise (sim quarter-dose)
Restoreconservative "do-no-harm" edits
Score4 layers of task-based evidence
Task rescue · CRM (N=50 OOD)
+0.0324
best target-gain vs 4 filters
External transfer · Mayo
+2.35 dB
PSNR, 95% CI 2.03–2.67, 100% win
Anatomy · Dice vs full-dose
0.61→0.74
edits land on real structures
Stability · Monte Carlo
62%
noise-sensitive (honestly disclosed)
Three surfaces, one product
USE

SafeRestore Studio

Conservative restoration that never ships a bare image — every output carries its evidence and a "needs review" flag. Revives older / low-dose scanners.

Owner: David · customer: imaging centers & scanner vendors
PROVE

Evidence API

Vendor-neutral. Score any restoration model on 4 layers → Evidence Score, failure-case audit, risk memo. The independent layer before pilots & procurement.

Owner: Andrew · customer: imaging-AI developers, PACS/scanner vendors
TEACH

SafeAI Academy

The same engine, reversed. Students learn why "looks sharper" isn't enough and spot where AI hallucinates — capstone runs Use + Prove live.

Owner: Nick · customer: med schools, residents, STEM education
Use · Studio  —  owner: David

Restoration you can trust, because it ships with proof

Upload a real 3D CT folder (DICOM series) and it is restored by the actual trained DeblurUNet25D model when the local server is running — otherwise an in-browser phantom keeps the demo working offline. The server sorts the slices, denoises the volume with 2.5D neighbor context, and shows a representative denoised slice with volume-level improvement metrics. Crucially, the image always comes bundled with its evidence.

checking backend… ready
Full-dose reference
Quarter-dose input
SafeRestore output
Input PSNR
Restored PSNR
Δ improvement
Max abs change

Evidence shipped with this image auto-generated

Run a restoration to generate the evidence panel…
Not for diagnosis. Phase 1 output is a retrospective, de-identified research/QA artifact. Every restored series requires radiologist review; flagged regions must not be interpreted from the enhanced image alone.
Prove · Evidence API  —  owner: Andrew

Score any restoration model — independently

Pick a model and run the validation protocol. The same 4 layers run on anyone's model; numbers below are the notebook's strict held-out runs (RSNA N=50 OOD, Mayo external, TotalSegmentator).

Evidence Score
select a model

Evidence Score — all methods on the same OOD set

Composite of task-rescue, win-rate, low-iatrogenic, and fidelity. The point isn't "sharpest" — it's the best benefit/safety trade-off on data the model never saw.

Deep Evidence Pack V-Ultimate

Failure-case audit exposes risk, not hides it

Executive risk memo

Run validation to generate the memo.
Ready for: research / QA / pilot evidence
Not for: clinical interpretation
Teach · Academy  —  owner: Nick

The same engine becomes a classroom

Degradation isn't only for training models — it's how you train people. Students learn the one lesson that powers the whole product: a sharper-looking image can still be wrong.

🔍 Lab: spot the hallucination

One of these AI outputs invented a structure that is not in the ground truth. Which one is unsafe?

Pick A, B, or C.
Score: 0 / 0

🎚 Lab: why "looks sharper" ≠ "is correct"

Drag to degrade the scan. Watch the downstream AI's confidence and the measured PSNR fall together — the eye is fooled long before the metric is.

AI detection confidence
0.86
Image PSNR

Teaching point: this is exactly why SafeRestore validates with task-based evidence, not PSNR/SSIM alone.

Capstone path — the CT work is the finale, not the start

1 How machines learn from examples · images as number matrices
2 Train / validation / test splits — why leakage lies to you
3 Probabilities, thresholds, precision & recall
4 AI safety: bias, overfitting, hallucination
Capstone — CT restoration safety: degrade a scan, restore it, then run the Evidence API and defend whether it's safe.