Planned research
SLZ Lumen
Planned research · Synthetic prototype reviewed · Clinical study planned
Study radiotherapy dose prediction and whether uncertainty identifies unacceptable errors better than anatomy-based risk at a fixed review workload. The initial population will follow suitable authorized planning data; the organ remains unselected.
Public summary updated
What the synthetic experiments found
The original toy network lost to a constant predictor. A controlled initialization experiment then reduced development mean absolute error to 0.363 versus 0.877 toy_Gy for the constant. Predictions remained flat inside every target, and target D2/V10 errors worsened versus unchanged initialization. This is not clinical planning accuracy.
Next question
Source-neutral input, reference-dose and geometry requirements are documented. Synthetic metadata-handling checks passed, but a separate review-interface qualification stopped at a layout boundary. A possible correction was identified through static review and remains untested; the cause of the failure is unknown. Next: verify source suitability and permissions, then obtain medical-physics review of anatomy, dose provenance and failure limits. No organ or source is selected; liver remains a conditional extension. Predicted research dose maps do not establish deliverable treatment plans.