基于生境信息的MRI放射组学与深度学习融合预测胶质母细胞瘤短期生存

RATIONALE AND OBJECTIVES: To develop and externally validate a preoperative multicontrast MRI stacking model integrating unsupervised habitat radiomics and deep learning to predict short-term survival (overall survival ≤ 9 months) in glioblastoma, and to explore transcriptomic and immune correlates.

MATERIALS AND METHODS: Retrospective multicenter study of 468 patients (training, n=196; external validation cohorts 1-3, n=66, 185, and 21). Radiomics models derived from whole-tumor, MRI-defined subregions, and unsupervised habitats were compared with a 3D ResNet-50 model. A two-level stacking ensemble combined habitat radiomics and ResNet-50 predictions. Performance and clinical utility were evaluated by area under the receiver operating characteristic curve (AUC), decision curve analysis, and related metrics. In external cohort 3 with paired RNA sequencing, patients were stratified by fusion-model predictions for exploratory differential expression and immune deconvolution, followed by Spearman correlation analyses.

RESULTS: Short-term survival occurred in 159/468 (34.0%). Habitat radiomics outperformed whole-tumor and MRI-defined subregion radiomics in external cohorts (AUC, 0.722-0.761), while 3D ResNet-50 achieved AUCs of 0.711-0.739. The fusion model achieved an AUC of 0.974 in the training cohort and AUCs of 0.860, 0.861, and 0.789 in external validation cohorts 1-3, respectively, and yielded the highest net benefit across clinically relevant thresholds. In the RNA-sequenced cohort, DLK1 expression was positively correlated with inferred T follicular helper cells and M1 macrophages.

CONCLUSION: Habitat-informed fusion of radiomics and deep learning enables preoperative prediction of short-term survival in glioblastoma. Radiogenomic analyses in the paired cohort are exploratory and suggest a DLK1-related immune-infiltration correlate of the model-defined phenotype.