Evaluating sequence contributions to MRI radiomics for glioblastoma survival: single vs fusion models

Glioblastoma is the most aggressive primary brain tumor with a poor prognosis. Multiparametric MRI-based radiomics shows promise for prognosis, yet most studies fuse all sequences without quantifying their individual contributions. We systematically compared the prognostic value of features from single sequences and comprehensive fusions for survival classification.

This retrospective study included glioblastoma patients from TCIA. Quantitative features were extracted from T1, T1-GD, T2, and T2-FLAIR images. Univariate ROC analyses with Benjamini-Hochberg false discovery rate correction assessed each feature's prognostic value. For multivariate modeling, combining univariate analysis with the maximum relevance minimum redundancy algorithm, was used to select the top predictors for each model. Logistic regression models were then built using features from single sequences, dual-sequence fusions, and a comprehensive four-sequence fusion. Performance was compared using repeated random subsampling validation with AUC as the metric.

Univariate analysis identified 249 features with significant prognostic power (T1-GD: 79; T2: 58; T2-FLAIR: 57; T1: 55). In the multivariate analysis, the four-sequence fusion model achieved the highest performance on the training cohort (AUC: 0.8467), but this advantage did not generalize to the validation cohorts. On the validation set, single-sequence models achieved AUCs ranging from 0.6599 to 0.7010, with the T2 model performing best. The dual-sequence model combining features from T1-GD and T2 sequences yielded the highest overall performance, achieving a mean validation AUC of 0.7066. Notably, this targeted two-sequence model outperformed the more complex four-sequence model (AUC: 0.7030).

For glioblastoma survival prediction, a strategic selection of complementary imaging sequences might be more effective than an indiscriminate aggregation of all available structural MRI data. However, this finding is currently specific to this particular dataset and clinical task. Future research using large-scale data with multiple tumor types and clinical objectives is necessary to explore the broader generalizability of these results.