CTV-AD Considerations for Future Editions
From QIBA Wiki
These items came up during development of the CTV-AD Profile publicly reviewed stage for publication in 2016.
They may be considered/addressed during future editions of the profile.
- We will not publish the Technically Confirmed Profile without replacing the dataset for the segmentation performance assessment. Otherwise we will have the problem of algorithms training on the test set. Consider adding at least a couple of single timepoint liver studies to segment multiple times. And/or get Ehsan to generate some known truth images.
- Consider if additional Periodic Calibration requirements are needed to (better) achieve the claim (Ehsan)
- Consider validating (in the field test or a groundwork project) the statement that "the profile constraints on voxel noise and spatial resolution are considered adequate to prevent any algorithm effects that would significantly impact the segmentation process and the resulting claimed performance. (Ehsan)
- Consider whether constraints on noise texture metrics (such as average frequency of the noise power spectrum in mm-1) would result in improved segmentation and Volumetry (Ehsan). Evidence is being investigated, and whether texture measured in phantoms will correlate to noise texture in patients for iterative algorithms.
- Consider whether it is likely that any "reasonable" protocol would cause INPLANE resolution problems that would degrade segmentation accuracy and repeatability (Kirsten; Mike, Ehsan, Mario). Can we relax the resolution requirement?
- Consider whether metrics on phantoms with more features would help; e.g. GGO conspicuity, air pocket, polystyrene, or low contrast spatial resolution for liver where iterative effects might be more of an issue? Although requiring specific phantoms or features makes conformance more difficult.
- Consider whether the noise/resolution assessment procedures should be modified to validate protocols across a wider range of patient characteristics
- Consider whether the results generated by different MTF software (e.g. vendor tools, TG233 reference tool, etc) differs enough to require validation (Like the 4.3 procedure)
- Consider whether TG233 (April 2016) came up with anything to review with vendors and include in the profile (Ehsan)
- Consider if 5HU consistency requirement is actually effective in preserving segmentation performance. It is tedious
- Expand the dataset for the Tumor Volume Change Repeatability Assessment Procedure. It covers tumors 10mm-67mm while the Profile goes up to 100mm. It is entirely lung data, while the profile aspires to cover tumors throughout the chest, abdomen and pelvis.
- Address the issue of systems/sites training on our (limited) test dataset.
- Consider allowing (requiring?) sites to do test-retest scans on their own for their assessment procedures and submit the data to QIDW so we can grow it. (Note the potential dose issues and need for easy QIDW submission)
- Appendix C: Between now and Tech Confirmed, take another run through the profile in general and this section specifically to keep/drop/update, and include a list of acronyms if keeping, or consider a QIBAWiki page that can be referenced instead.