Deep learning prediction of overall survival in glioblastoma patients using pre-chemoradiotherapy MRI and clinical data

Abstract

Glioblastoma (GBM) exhibits variable treatment response and patients shown a median survival of 15 months. A major challenge lies in the difficulty of reliably assessing treatment response at an early stage, as it can usually only be evaluated on anatomical MRI performed up to three months after the start of treatment, when early interpretation may be complicated by the risk of pseudoprogression. Accurately predicting short-term and long-term survivors before radiotherapy and chemotherapy is therefore crucial. The objective of this study is to develop an AI model designed to predict survival outcomes based exclusively on pre-chemoradiotherapy MRI and clinical data. A total of 1,575 retrospective T1-Gd brain images were obtained from 175 patients treated for GBM between 2018 and 2023. A ResNet-51Q model pretrained on ImageNet was fine-tuned on pre-chemoradiotherapy MRI using transfer learning. To incorporate clinical data, deep hybrid learning (DHL) models were trained. At patient-level, the AI model achieved 81.08% (CI 95%: 67.57–91.89), 68.42% (CI 95%: 47.06–88.89) and 94.44% (CI 95%: 81.82–1.00) for accuracy, sensitivity, and specificity, respectively, for the classification of short-term and long-term survivors using only pre-chemoradiotherapy MRI. At patient-level, the three-class model using only pre-chemoradiotherapy MRI, achieved 87.18% (CI 95%: 76.92– 97.44), 87.18% (CI 95%: 76.68–96.97), and 93.59% (CI 95%: 88.03–98.41) for accuracy, sensitivity, and specificity, respectively. In comparison at slice-level, the three-class model using only pre-chemoradiotherapy MRI, achieved 81.77% (CI 95%: 70.66–91.45), 81.77% (CI 95%: 71.07–91.33), and 90.88% (CI 95%: 85.25–95.74) for accuracy, sensitivity, and specificity, respectively. In comparison the DHL model with clinical data achieved an accuracy of 84.62% (CI 95%: 71.79–94.87), a sensitivity of 84.62% (CI 95%: 72.54–95.24) and a specificity of 92.31% (CI 95%: 86.07–97.53) and 76.32% (CI 95%: 63.16–89.47), 76.11% (CI 95%: 61.63–88.75) and 92.06% (CI 95%: 87.51–96.23) for four overall survival classification. For multiclass tasks, metrics were calculated using a macro-averaged approach. The combination of pre-chemoradiotherapy MRI and clinical data using the DHL model for early survival prediction; however, adding clinical variables yielded comparable prognostic performance rather than a demonstrable enhancement over imaging features alone. The use of AI models to predict survival outcome in GBM from pre-chemoradiotherapy data may support pre-chemoradiotherapy prognostic stratification.