Evidence
A library of evidence, not a single score.
Choose the question you want to investigate. Each area offers a readable explanation and the corresponding existing machine-readable report.
Imaging evidence at a glance
Each card names its source, what was measured and how to interpret it. These recorded results come from separate activities. Harmonisation describes a future SPE evaluation.
Engineering measurements describe the tests that ran. Official evaluation requires approved data definitions and a locked protocol.
Explore neural network evidence →Clinical baseline
Understand the artificial reference and its stored distribution, balancing and reconstruction measurements.
Open this activityReconstruction quality
Inspect per-variable errors, class recall and single-versus-multiple masking.
Open this activityNeural network studies
Follow the architecture, learning curve, comparisons, unsuccessful branches and resource evidence.
Open this activityCT benchmarks
Compare local structure, intensity error, body-envelope results and HU bands.
Open this activityModels and governance
Distinguish the deployed model from specialist experiments and understand human approval gates.
Open this activityReadiness and next steps
Separate implemented infrastructure from official evaluation prerequisites.
Open this activityTechnical imaging evidence bundle
Read the stored public-example and artificial-phantom reports with their source fingerprints. The readiness and neural pages retain their own complete contracts and report downloads. Opening these stored files does not run an analysis.
Terms explained
Technical names remain available for reviewers, but you do not need them to begin an activity.
Open the plain-language glossary
- HU — Hounsfield units
- The CT intensity scale. Image similarity and intensity accuracy are different questions.
- kVp
- Peak X-ray tube voltage, one acquisition setting. It is not a complete radiation-dose measurement.
- MAE and RMSE
- Average absolute error and root mean square error. Both measure numerical differences; RMSE penalises larger errors more.
- SSIM
- Structural similarity. Local masked SSIM focuses on the hidden region; global SSIM can be dominated by unchanged pixels.
- PSNR
- Peak signal-to-noise ratio, an error-based similarity measure. Higher is better within the same scaling and evaluation settings; it is not clinical certification.
- W₁ / Wasserstein distance
- A difference between intensity distributions. It does not tell you whether structures are in the correct location.
- GLCM Joint Entropy
- A texture measure derived from how neighbouring intensity levels occur together.
- CCC
- Concordance correlation coefficient: an agreement measure. Robustness is assessed separately at each declared noise level.
- ICR
- Imaging completion rate: accepted complete acquisitions divided by predefined eligible missing-acquisition tasks. It is not a slice reconstruction score.
- KS distance
- A difference between cumulative distributions. Lower means closer distributions, not guaranteed individual accuracy.
- Class recall and balanced accuracy
- Recall is the fraction recovered within a true class. Balanced accuracy averages recalls so a frequent class does not dominate.
- Macro-F1
- An equal-class average combining precision and recall. It complements, rather than replaces, the class confusion matrix.
- DCR / NNDR
- Distance to the closest record and nearest-neighbour distance ratio. These are similarity-risk diagnostics, not anonymity guarantees.
- Small-cell suppression
- Groups below the declared support threshold are withheld or marked insufficient rather than assigned unsupported precision.
- Condition-only generation
- A proposed route from an approved synthetic clinical profile and stochastic anatomy to new CT acquisitions. This is distinct from reconstructing a slice of an existing study.
- SPE and fail-closed gates
- The authorised Secure Processing Environment and checks that prevent protected execution when required data definitions or approvals are missing.
Existing OpenAPI schema → — availability follows the deployment's current access policy.