A Deep Learning Approach to Glioblastoma Radiogenomic Classification Using Brain MRI

MGMT promoter methylation prediction based on multiparametric MRI via vision graph neural network

“…The unique design of skip connections solved the over-smoothing problem, which limits the potential of deep GCNs. Validated on the RSNA radiogenomic dataset, our model outperformed the state-of-the-art methods [25][26][27]47 that won the challenge with an averaged validation and test AUC score of 0.653 and 0.628, respectively, as shown in Table 3 and Fig. 7.…”

“…In this section, we evaluate the effectiveness of the multi-sequence fusion strategy and assess the statistical significance of our ViG model’s performance in the MGMT prediction. We compare our approach with several winning models trained on single MR sequences each time 27 . The original authors only reported cross-validation results in their paper.…”

“…We compare our approach with several winning models trained on single MR sequences each time. 27 The original authors only reported cross-validation results in their paper. We extended their model to fused multisequence MRI.…”

“…The key idea behind Bayesian optimization is to constrain the hyperparameter search space to reduce the optimization time. Emchinov 27 trained a ResNet-34 model for each magnetic resonance imaging (MRI) modality and then ensemble the models by averaging their outputs to produce the final prediction probability. Although these CNN-based methods performed well in the competition, the use of the convolution operation with sliding windows in these models 28 introduced locality and inflexibility when handling these images, resulting in the loss of important global information.…”