Overall Survival Prediction in Glioblastoma Using Clinical Factors Combined with Texture Features Extracted from 3D Convolutional Neural Networks

Background: Glioblastoma (GBM) is the most common primary brain tumor and has a poor prognosis. Accurate overall survival (OS) prediction may allow for personalized treatment recommendations. Objectives: We aimed to predict OS in GBM patients following gross total resection (GTR) using preoperative MRI images. Methods: A cohort of 87 GBM patients (59 patients for training and 28 patients for validation) who underwent GTR was analyzed using multi-institutional data from the 2018 Brain Tumor Segmentation (BraTS) Challenge [1,2]. Each patient dataset consisted of a series of preoperative MR images including T1, T1 with contrast, T2, and T2-FLAIR images. A group of experienced radiologists delineated areas of tumor core, tumor enhancement, and surrounding edema for each patient dataset using these image sequences. A 2D U-Net was trained to segment these structures on the validation cohort. A 3D convolutional neural network model with orthogonalized random filters was used to learn image features from the three segmented subregions including texture, size, location, etc. Global maximum pooling was performed on intermediate convolutional layers to obtain representative image features for each patient [3]. MR images from both short- and long-term survivors were augmented with random rotations to balance the number of patients among three cohorts of patients, as midterm survivors (6-18 months) outnumbered short-term (\u3c6 months) and long-term (\u3e18 months) survivors by a large margin The extracted image features were then fed into a radial basis function (RBF) kernel-based L-2 norm regression algorithm [4] to predict OS. Results: The average [standard deviation] Dice similarity coefficient for the whole tumor, enhanced tumor, and tumor core contours were 0.882 [0.080], 0.712 [0.294], and 0.769 [0.263], respectively for the validation cohort. Parameters of regression algorithm were optimized using leave-one-out cross validation. One convolutional layer was used in the CNN architecture due to limited training samples. Model performance deteriorated when using deeper layers. The best architecture to classify patients into short-, mid- and long-term survivors was one convolutional layer with 30 filters. The overall prediction accuracy was 64.3% and the Spearman correlation was 0.513, achieved by assembling predictions made from filters of size 4×4×3 and 5×5×3. Model performance was further improved by including clinical factors such as age, tumor location, and volumetric tumor-to-brain ratio; and accuracy improved to 67.9% while the Spearman correlation was 0.503. Experimental results have further verified age as an important index in OS prediction. Conclusions: We developed a 3D multi-scale CNN model with kernel regression to extract image features from preoperative MR images and predicted OS in GBM patients with 67.9% accuracy. Further studies will explore additional image features and biomarkers