Multiscale MRI deep learning predicts short-term survival in glioblastoma

Multiscale MRI deep learning predicted <=9-month survival in glioblastoma with external validation discrimination values of 0.871, 0.828, and 0.798; gains over clinical-MRI baseline were 0.161 and 0.131, with only cohort 1 significant after false discovery rate correction.

• Model input integrated whole-brain, three-dimensional tumor, and 2.5-dimensional tumor MRI. It was trained on 290 adults and externally tested on 225, 182, and 31 adults with newly diagnosed glioblastoma.
• Compared with a combined clinical-MRI morphometric baseline, the deep learning model improved area under the receiver operating characteristic curve (AUC) by 0.161 in external cohort 1 and 0.131 in external cohort 2; only cohort 1 remained significant after false discovery rate correction.
• Model output was linked to immune/inflammatory and cell-division/genome-maintenance transcriptomic pathways, concordant and false discovery rate-significant in both cohorts.

Evidence grade: Level 3

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.