Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

Survival Prediction in Glioblastoma Using Combination of Deep Image Features, Hand-Crafted Image Features and Clinical Factors

Y Zhuge*, H Ning, J Cheng, P Mathen, K Camphausen, R Miller, A Krauze, National Cancer Institute, Bethesda, MD


PO-GePV-M-43 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Glioblastoma (GBM) is the most common malignant primary tumor of the brain. Overall survival prediction in GBM patients plays an important role to determine the most appropriate treatment plan, and to facilitate preoperative care of patient. In this study we propose a novel method for overall survival prediction in GBM by using combination of deep MRI image features, hand-crafted image features, and clinical factors.

Methods: The proposed method consists of three steps: (1) 3D brain tumor segmentation based on the popular U-Net model; (2) Patients are then classified into three groups as long survivors (e.g., >900 days), short survivors (e.g., <300 days), and mid-survivors (between 300 and 900 days) using the Dense-Net model on segmented brain tumor regions. Deep MRI image features are extracted through the last fully connected layer of the classification model. Hand crafted image features, including first-order statistics, shape features, and texture features are extracted by using pyRadiomics on MRI images; (3) A Cox proportional hazards neural network named DeepSurv is utilized for survival prediction on the features combining the deep image features, hand crafted image features, and clinical characteristics.

Results: The proposed scheme are evaluated on the Multimodal Brain Tumor Image Segmentation (BraTS) Benchmark 2020 training data sets. All data are divided into training (60%), validation (20%), and testing sets (20%). The Concordance Index (C-index) of the training, validation, and testing sets are 0.985, 0.878, and 0.846, respectively. The proposed method is compared with the random survival forest approach which is a traditional machine learning-based prediction method. The C-index values on the same data sets are 0.797, 0.842, and 0.805, respectively.

Conclusion: The method has the potential to be an imaging biomarker for prediction of the overall survival in patients with GBM; and has the potential to facilitate the preoperative care of patients with GBM.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email