Prediction of Overall Survival Time in Glioblastoma Patients Using Deep Neural Networks

Introduction: Glioblastoma, the most aggressive primary brain tumor, presents a mean overall survival of approximately one year, posing significant challenges for effective treatment planning. Previous machine learning studies for glioblastoma survival prediction have achieved limited accuracy due to small sample sizes and have poor scalability due to reliance on time-consuming manual image segmentation. This study aims to develop a novel ML model for automatic tumor segmentation and accurate prediction of survival for glioblastoma using a large set of MRIs for treatment decision-making.

Methods: MRI images from 225 patients in the multicenter open-access BRATS dataset and 18 Johns Hopkins patients were used to train the model. An Unet with Transformer model was developed to automatically segment the necrotic tumor core, enhanced tumor region, and surrounding edema. For survival prediction, an 18-layer ResNet3D model (which used MRI images and the segmentations as inputs) was developed. It had four residual blocks containing convolutional layers to ensure efficient feature extraction; batch normalization, ReLU activation, and skip connections were added to prevent vanishing gradients. The performance for segmentation was evaluated using the Dice similarity score (measuring overlap between the segmentations and ground truth), and the accuracy of survival prediction was evaluated with mean squared error.

Results: The segmentation model achieved a Dice score of 0.899 on the whole tumor region, suggesting a strong ability to distinguish tumor regions from other structures. For survival prediction, the ResNet model achieved a training MSE of 3775 (error of 2.04 months). On testing (unseen) data, the model achieved an MSE of 82103 (error of 9.55 months).

Conclusion : Our models generate more accurate segmentations and survival predictions than prior literature using the BraTS dataset. This provides significant promise for the ability to more accurately estimate patient prognosis to guide clinical management.