Radiomic AI Models Show Sensitivity to CT Acquisition Parameters
A new study from arXiv (2605.14667) quantifies how variations in CT scan acquisition parameters degrade the performance of radiomic AI models, a key barrier to clinical deployment. The researchers developed a mixed-effects framework to identify clinically significant parameter regions that improve cross-dataset robustness. Applied to lung cancer diagnosis using two independent multicenter datasets (a public database and own-collected data) with several state-of-the-art architectures, the framework revealed that adjusting CT parameters based on collected data enhanced reproducibility when tested on the public set. The optimal configuration is currently being validated.
Key facts
- Study quantifies scan parameter sensitivity of radiomic AI models
- Mixed-effects framework accounts for subject-level random effects
- Applied to lung cancer diagnosis in CT scans
- Two independent multicenter datasets used: public database and own-collected data
- Several state-of-the-art architectures tested
- CT parameters adjusted using collected data and tested on public set
- Optimal configuration currently being validated
- Published on arXiv with ID 2605.14667
Entities
Institutions
- arXiv