Predicting the survival of patients with glioblastoma using deep learning: a systematic review

Abstract

High-grade gliomas (HGG) are aggressive brain tumor and despite intensive treatment strategy have a relatively low rate of overall survival (OS). There is no reliable technique for prediction of survival of HGG. We aimed to investigate the performance of deep learning (DL) algorithms for predicting OS of patients with glioblastoma. A systematic search was conducted in well-established research databases from inception to 23 May 2023 to retrieving the eligible studies. The sensitivity, specificity, and accuracy regarding DL algorithms regarding OS of glioblastoma was extracted. A total of 19 studies were included: 11 used conventional neural networks (CNNs) and eight used support vector machines (SVM). 17 studies performed validation, with 16 using cross-validation or Leave-One-Out Cross-validation. The radiomics features extracted varied from 3 to 17,441. Transfer learning was used in 6 studies. Several studies evaluated the accuracy, sensitivity, specificity, and AUC of DL models for GBM survival prediction. The accuracy ranged from 46.4 to 98.4% for CNNs to SVMs models. Sensitivity varied from 42.9 to 96.5%, while specificity ranged from 16.7 to 99.0%. The AUC values ranged from 61.4 to 85% for CNNs and SVMs models. Depending on multiparametric imaging data, DL can help with glioblastoma patient stratification, but external multicenter repeatability studies are needed before therapeutic integration. Radiomics-guided precision oncology shows promise for optimizing glioblastoma care. Future research should focus on developing predictive models that incorporate larger patient cohorts and more robust imaging modalities. Dataset harmonization is also necessary for effective risk categorization.