Studio
You do not need to know the model.
A plain-language route through every activity, result and limitation in the demonstrator.
This tour explains what each activity does, what to try and what the result means. The clinical examples are artificial. CT previews come from a licensed public research example. You cannot upload patient data here.
Start with the reference
The cohort is the comparison point, not a collection of real patients. Open its composition and missingness. The stored benchmark sits at the bottom of Generate, below the current result. You will see why balancing a group and preserving every distribution are different objectives.
Open this part of the studio →Create the group you requested
Request 40 records with a 70% female share, then generate. Read requested, obtained and shortage first. An achieved quota can coexist with an unmet generic imbalance threshold. Try reference-distribution mode to ask a different question. Advanced options expose stage, histotype, seed and the distance filter.
Open this part of the studio →Hide information you can check
Choose an artificial example and one to six fields. The service hides those values, reconstructs them and compares against the originals. A mismatch is an observed prediction error, not necessarily a broken workflow. Fields you did not hide must remain unchanged.
Open this part of the studio →See where predictions struggle
Choose a variable. In a confusion matrix, rows are true classes and columns are predictions; the diagonal counts correct answers. Recall shows the fraction recovered in each true class. Numerical errors are split by range. Compare single and multiple masking with their sample counts, not just the delta.
Open this part of the studio →Know the imaging source
Read the collection, licence and inventory. Hashed keys link the same series across pages. kVp identifies an acquisition setting, not a dose certificate. Our series have no available kVp metadata. Domain comparisons await a suitable SPE sample; the example demonstrates inventory, geometry and reconstruction.
Open this part of the studio →Hide part of a CT image
Choose a slice, method and mask, then reconstruct. Compare original, hidden input, reconstruction and error map; use the overlay to inspect differences. Central masking leaves visible pixels outside the mask, whose error should be checked separately. Structural similarity does not imply quantitative HU fidelity.
Open this part of the studio →Look beyond a favourable image
Load the stored all-slice benchmark. Read medians, quartiles and the worst values for local SSIM and intensity errors. SSIM ≥ 0.85 is a descriptive threshold only. The body envelope excludes much background; HU bands are not segmented organs.
Open this part of the studio →Understand the neural network contribution
Follow the network diagram, compare against linear and cubic interpolation, and inspect the 2-to-16-subject learning curve. The history includes unsuccessful branches. These are same-collection masked-slice studies, not evidence of complete CT-acquisition generation.
Open this part of the studio →Inspect harmonisation and robustness
Start with the prepared 29-field SPE adapter in the combined readiness page. kVp comparisons await an authorised SPE sample. Read the stored phantom results to understand the measurement workflow. Phantom robustness tests noise at 5, 10 and 20 HU separately; passing one level does not pass the others.
Open this part of the studio →Distinguish deployed AI from experiments
Read the live model card and the separate age/PD-L1 studies. The VAE KL coefficient called beta is not a Beta distribution. Experimental success does not mean a component is deployed or clinically validated.
Open this part of the studio →Finish with the remaining prerequisites
Readiness comes last so you can first see what works. A harness can be implemented while authorised data, reference tools or approvals remain pending. “ICR not evaluated” means no acquisition-completion evaluation was performed, not a measured 0% completion rate.
Open this part of the studio →Keep the evidence you need
The five “Imaging evidence at a glance” cards appear automatically with sources and explanations. Follow the neural network link for the learning curve; use aggregate JSON for technical review. Activity downloads label when they contain artificial demo records. The scope and provenance travel with the result. Interactive reconstruction does not update the fixed Variable diagnostics benchmark.
Open this part of the studio →Three things that should not be confused
Working activity: the interface completes the requested operation and reports its outputs. Observed limitation: a prediction or benchmark does not meet a target. Missing prerequisite: a protected evaluation is intentionally not run. A limitation or expected hold is not, by itself, an operational defect.