P13.10.a predicting survival of glioblastomaafter radiotherapy using deep learning and neuroimaging in a multi-centre cohort

Abstract BACKGROUND Glioblastoma is an aggressive brain tumour routinely monitored with MRI. However, the potential to use imaging as a prognostic biomarker predicting survival after treatment starts is unclear. This study aims to predict survival within eight months of completing radiotherapy from MRIs using deep learning. MATERIAL AND METHODS The dataset consists of 206 prospective and retrospective consecutive patients with glioblastoma (WHO 2021) across 11 UK centres (short-term survival: N=64; 31.1%). Models were trained/validated on N=158 (76.7%) retrospective patients (3 centres). Two holdout sets were sampled: retrospective test set (N=19); and external, prospective test set (N=29; 8 separate centres). Dense convolutional neural networks were developed with separate branches for T1c-w and T2-w MR sequences; a third branch concatenates outputs from those branches to predict survival. Dense blocks used pretrained weights for an abnormality detection model. A semi-supervised approach was implemented; patients without known outcomes were added with pseudo-labels during training. Parameter tuning included the number of updated blocks, linear layer sizes, and weighting of pseudo-labels. Imaging models were compared to counterparts that apply non-imaging variables (demographics, MGMT status/%, treatment). Logistic regression, support vector classifier, and decision tree models were fit on non-imaging features alone with feature selection and mean/mode imputation. Combined imaging/non-imaging networks were developed that added a branch for non-imaging inputs. RESULTS The current best-performing imaging model used MR sequences alone (validation weighted-ROC=0.80, 95% CI=0.75-0.84; balanced accuracy=0.70, 95% CI=0.55-0.85). Among non-imaging models, the logistic regression model with feature selection performed best (validation weighted-ROC=0.59, 95% CI=0.56-0.63; balanced accuracy=0.59, 95% CI=0.56-0.63). Combined model development is ongoing. CONCLUSION This is the first known model applying deep learning to MRIs from multiple centres to distinguish patients with failure to respond to chemoradiotherapy, and those who survive the subsequent treatment window. Based on initial results, image-based models performed best at predicting survival. Such models could be used across centres to suggest closer monitoring and trial targeting of patients with expected short-term survival compared to treatment response.