Computed tomography (CT)
Harmonisation, Robustness & Readiness
Working software, an achieved engineering threshold and authorisation for official evaluation are different states. This page explains the prerequisites without treating an expected hold as a broken feature.
The connection to official data is prepared
The Secure Processing Environment (SPE) adapter defines 29 clinical fields and explicit source-to-model mappings. It connects authorised clinical data with study and series metadata. Dictionary checks, hashes, missing/unknown states, subject-level splits and separate training approval govern protected execution. Public templates await confirmation from the data owner.
SPE compatibility and controlled delivery path
The 29-field adapter is a preparation layer. Templates remain unresolved until authoritative semantics, linkage, splits and separate training approval are supplied.
Robustness: a prepared measurement workflow
The recorded test adds Gaussian noise at 5, 10 and 20 Hounsfield units (HU) to nine artificial phantom series. It measures agreement of a texture feature before and after noise. Each level is assessed separately, revealing where the workflow needs further development.
Official CT evaluation requires an authorised sample and approved lung segmentation and extraction settings. This measurement workflow is not a regulatory certification.
Evaluation tools prepared for future official testing
The readiness report covers its declared KPI scope. It must not be read as a score for all four official KPIs; clinical quota control alone does not complete KPI 2.
Harmonisation evaluation — awaiting SPE data
Harmonisation aims to make scans from different protocols more comparable while preserving meaningful image features. No CT series currently available to us has kVp metadata. Acquisition-domain evaluation awaits an authorised SPE sample with suitable metadata and comparable protocols.
That evaluation will compare Wasserstein distance before and after mapping, together with stability of GLCM Joint Entropy, a texture feature. Evaluation on real CT data has not yet been performed.
Earlier software test on artificial images
Nine computer-generated CT image series (phantoms) simulated three acquisition settings. A statistical mapping adjusted their Hounsfield unit (HU) intensity values towards a common reference. On separate evaluation samples, the median Wasserstein distance between the three pairs of intensity distributions fell by 99.3%.
This demonstrates distribution alignment in a controlled artificial test; it does not establish preservation of anatomy or radiomic features. The separate Robustness test above asks a different question: whether a radiomic feature stays stable when noise is added. Validation of harmonisation on real CT scans awaits suitable SPE data.