Multiscale Multiparametric MRI Deep Learning for Short-Term Survival Assessment in Glioblastoma.

Background

Preoperative identification of short-term survival in glioblastoma may guide management but remains challenging because of clinical and imaging heterogeneity.

Purpose

To develop and externally validate a multiscale magnetic resonance imaging (MRI)-based deep learning model for short-term survival assessment and explore transcriptomic correlates. STUDY TYPE Retrospective, multicenter. POPULATION Adults with pathologically confirmed, newly diagnosed glioblastoma (n = 728): training cohort (n = 290; median age, 55 years; 169 men) and external cohorts 1-3 (n = 225/182/31; median ages, 64/61/56 years; 136/108/20 men, respectively). FIELD STRENGTH/SEQUENCE 1.5 T or 3.0 T; axial precontrast T1-weighted, T2-weighted, T2-weighted fluid-attenuated inversion recovery, and postcontrast T1-weighted MRI. ASSESSMENT Short-term survival was overall survival of 9 months or less. Whole-brain, three-dimensional tumor, and 2.5-dimensional tumor inputs were integrated and compared with clinical, conventional MRI morphometric, and combined clinical-MRI baselines. STATISTICAL TESTS Kruskal-Wallis, Mann-Whitney U/Wilcoxon rank-sum, chi-square, Fisher exact, DeLong, and log-rank tests; Benjamini-Hochberg false discovery rate (FDR) correction; calibration, decision curves, edgeR, and correlation-adjusted mean-rank gene-set testing were used. Areas under the receiver operating characteristic curve (AUCs) with 95% confidence intervals (CIs) summarized discrimination; threshold metrics used the Youden index. Two-sided p < 0.05 or FDR-adjusted p < 0.05 indicated significance.

Results

Apparent training AUC was 0.870 (95% CI: 0.830, 0.910); external AUCs were 0.871 (95% CI: 0.821, 0.920), 0.828 (95% CI: 0.761, 0.895), and 0.798 (95% CI: 0.640, 0.956). AUC gains over the combined clinical-MRI morphometric baseline were 0.161 and 0.131 in external cohorts 1 and 2; only cohort 1 remained significant after FDR correction (cohort 2, FDR-adjusted p = 0.0897). Immune/inflammatory and cell-division/genome-maintenance pathway associations were directionally concordant, significant after FDR correction in both cohorts, and leave-one-out consistent. DATA

Conclusion

Multiscale MRI deep learning demonstrated favorable discrimination for short-term survival; model output was associated with immune- and cell-cycle-related transcriptomic programs. EVIDENCE LEVEL 3. TECHNICAL EFFICACY Stage 2.