Glioblastoma and radiotherapy: A multicenter AI study for Survival Predictions from MRI (GRASP study)

Abstract

Background: The aim was to predict survival of glioblastoma at 8 months after radiotherapy (a period allowing for completing a typical course of adjuvant temozolomide), by applying deep learning to the first brain MRI after radiotherapy completion.

Methods: Retrospective and prospective data were collected from 206 consecutive glioblastoma, isocitrate dehydrogenase -wildtype patients diagnosed between March 2014 and February 2022 across 11 UK centers. Models were trained on 158 retrospective patients from 3 centers. Holdout test sets were retrospective (n = 19; internal validation), and prospective (n = 29; external validation from 8 distinct centers). Neural network branches for T2-weighted and contrast-enhanced T1-weighted inputs were concatenated to predict survival. A nonimaging branch (demographics/MGMT/treatment data) was also combined with the imaging model. We investigated the influence of individual MR sequences; nonimaging features; and weighted dense blocks pretrained for abnormality detection.

Results: The imaging model outperformed the nonimaging model in all test sets (area under the receiver-operating characteristic curve, AUC P = .038) and performed similarly to a combined imaging/nonimaging model (P > .05). Imaging, nonimaging, and combined models applied to amalgamated test sets gave AUCs of 0.93, 0.79, and 0.91. Initializing the imaging model with pretrained weights from 10 000s of brain MRIs improved performance considerably (amalgamated test sets without pretraining 0.64; P = .003).

Conclusions: A deep learning model using MRI images after radiotherapy reliably and accurately determined survival of glioblastoma. The model serves as a prognostic biomarker identifying patients who will not survive beyond a typical course of adjuvant temozolomide, thereby stratifying patients into those who might require early second-line or clinical trial treatment.

Keywords: artificial intelligence; deep learning; glioblastoma; magnetic resonance imaging; survival.

Figure 1.

Architectures for dense neural networks. (a) Imaging model: The model inputs whole brain contrast-enhanced T₁-weighted sequences, and T₂-weighted sequences as separate branches (T1c and T2 branches). These are passed through dense blocks with pretrained weights. Outputs are flattened and reduced before feature concatenation. Predictions are obtained from the merged linear layer (concatenating vectors from T1c and T2 branches). (b) Combined model: Modified version of the architecture with an additional branch consisting of nonimaging inputs and linear layers. For illustrative purposes, 3D MR volumes are shown as 2D images and 4D dense blocks as 3D representations.

Figure 2.

Receiver-operating characteristic curves for imaging, combined, and nonimaging models on holdout test data. (a) Model performances on the amalgamated test set. AUCs were 0.93, 0.91, and 0.79 for the imaging, combined, and nonimaging models respectively. (b) Model performances on the retrospective test set. AUCs were 0.92, 0.94, and 0.76 for the imaging, combined, and nonimaging models respectively. (c) Model performances on the external, prospective test set. AUCs were 0.93, 0.89, and 0.78 for the imaging, combined, and nonimaging models respectively. AUC: area under the receiver-operating characteristic curves.

Figure 3.

Receiver-operating characteristic curves displaying imaging model performances for additional analyses run on the amalgamated test set. (a) Permutation test results (full imaging model, AUC = 0.93; permutation test, AUC = 0.49*). (b) Results from ablation studies (full imaging model, AUC = 0.93; predictions from T1c branch, AUC = 0.83*; predictions from T2 branch, AUC = 0.85; trained model initializing random weights—ie, with no pretraining, AUC = 0.64*). Panels (c–f) show imaging model results disaggregated for sample subgroups. (c) Performances based on the initial surgery type (biopsy-alone, AUC = 0.89; resection, AUC = 0.87). (d) Curves plotted separately for age at first diagnosis (>60 y, AUC = 0.98; ≤60 y, AUC = 0.89). (e) Performances based on sex (female, AUC = 0.96; male = 0.89). (f) Performances split by the acquisition dimension of the input T1c MRI (2D, AUC = 0.90; 3D, AUC = 0.98). AUC: area under the receiver-operating characteristic curves. T1c: contrast-enhanced T1-weighted MRI. T2: T2-weighted MRI. *: significantly different AUC compared to the full imaging model using DeLong’s test with a threshold of P ≤ .05.

Figure 4.

Saliency maps from guided backpropagation on the merged branch of imaging models using T1c and T2 inputs. Patients from retrospective and prospective test sets were selected including erroneous classification predictions (patients 5 and 6). T1c: contrast-enhanced T₁—weighted MR sequence. T2: T₂—weighted MR sequence.