Predicting Survival of Brain Tumor Patients using Deep Learning
Sharmila Agnal A1, Arun Deepak C2, Venkatesh J3, Sudarshan S4, Pranav A5

1Sharmila Agnal A, Assistant Professor, Department of Computer Science and Engineering, SRM Institute of Science and Technology, Chennai (Tamil Nadu), India.
2Arun Deepak C, Department of Computer Science and Engineering, SRM Institute of Science and Technology, Chennai (Tamil Nadu), India.
3Venkatesh J, Department of Computer Science and Engineering, SRM Institute of Science and Technology, Chennai (Tamil Nadu), India.
4Sudarshan S, Department of Computer Science and Engineering, SRM Institute of Science and Technology, Chennai (Tamil Nadu), India.
5Pranav A, Department of Computer Science and Engineering, SRM Institute of Science and Technology, Chennai (Tamil Nadu), India.
Manuscript received on 07 April 2019 | Revised Manuscript received on 20 April 2019 | Manuscript published on 30 April 2019 | PP: 1441-1448 | Volume-8 Issue-6, April 2019 | Retrieval Number: F3652048619/19©BEIESP
Open Access | Ethics and Policies | Cite | Mendeley | Indexing and Abstracting
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open access article under the CC-BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)

Abstract: Deep Learning and advents in the field of machine learning, in general, has given rise to powerful classification algorithms applied to various real-world applications, hitherto requiring human experts. Consequently, the proposed system aims to automatically predict the survival of patients suffering from Glioma, a type of highly fatal brain tumor characterized by survival rates lower than two years. Until now, once a patient is diagnosed with Glioma it is the physician who provides the estimated number of days the subject would survive. However, the proposed system aims to automate the very process in order to obtain an unbiased prediction, bereft of any human error. Based on data consisting of Magnetic resonance imaging (MRI) images (four structural modalities), ground truth segmentation labels marking the region of interest (ROI) and accompanying information such as age and resection status of each patient, the system is expected to perform a classification task- high risk(less than 10 months survival), moderate risk(between 10 to 15 months survival) and low risk(greater than 15 months). While MRI image acquisition process can be complex, BraTS 2018 dataset fulfilled this vital requirement of the proposed system. The task entails two essential modules- feature extraction and classification. VGG16, a pre-existing Convolutional Neural Network (CNN) model, is employed to extract essential features from MRI images. Finally, an Artificial Neural Network (ANN) classifies the survival of a patient into the categories mentioned above. Prediction accuracy is measured using 5-fold cross-validation and a test dataset. The system achieved 100 % accuracy on training set itself, 88.8% using 5-fold cross-validation and 52% on an unknown test set, thereby, averaging 80.3% accuracy. While Glioma is, unfortunately, an extremely fatal condition, an accurate survival estimate can be a boon to the subject so as to fulfill one’s commitments and bid adieu to their loved ones. Thus, unbiased opinion from a well-trained prediction model can be greatly applied in medical institutions to give their patients the correct estimate of their survival.
Keyword: Deep Learning, Machine Learning, Glioma, MRI, VGG16, CNN, ANN, 5-Fold Cross-Validation.
Scope of the Article: Deep Learning