Influence of feature aggregation and selection methods on fluorine-18 fluorodeoxyglucose PET radiomics for survival prediction in patients with lymphoma.

OBJECTIVE

To investigate the influence of different feature aggregation and selection methods on the predictive performance of fluorine-18 fluorodeoxyglucose ( 18 F-FDG) PET radiomics in assessing survival outcomes in patients with lymphoma.

METHODS

This retrospective analysis included 80 patients with histologically confirmed lymphoma, each presenting with at least three lesions on baseline 18 F-FDG PET images. Metabolic tumor volumes were segmented using a standardized uptake value threshold of 4.0. From each lesion, 107 radiomic features were extracted. Of these, 30 features were preselected based on their robustness to variations in tracer uptake time, image reconstruction parameters, and respiratory motion. Six distinct feature aggregation approaches were evaluated in combination with six feature selection methods. Multivariable Cox proportional hazards regression was used to assess the predictive performance of each aggregation-selection strategy for progression-free survival (PFS) and overall survival (OS).

RESULTS

All combinations of feature aggregation and selection methods produced statistically significant prognostic models for PFS and OS, with Harrell's concordance indices (C-index) ranging from 0.582 to 0.668 for PFS and from 0.597 to 0.721 for OS. The best predictive performance was achieved using median value aggregation across all individual lesions combined with feature selection via the least absolute shrinkage and selection operator. Integrating clinical variables with radiomic features further improved predictive performance.

CONCLUSION

The prognostic value of 18 F-FDG PET radiomics remained consistent across different feature aggregation and selection strategies. The establishment of standardized analysis workflows is essential to facilitate its clinical implementation in personalized treatment planning for patients with lymphoma.