Machine Learning and Deep Learning Techniques to Predict Overall
Survival of Brain Tumor Patients using MRI Images [1]
Lina Chato
Department of Electrical and Computer Engineering University of Nevada, Las Vegas (UNLV) Las Vegas, Nevada
Lina.chato@unlv.edu
Shahram Latifi
Department of Electrical and Computer Engineering University of Nevada, Las Vegas (UNLV) Las Vegas, Nevada Shahram.latifi@unlv.edu
Summary of Results
This paper, published in 2017 for the IEEE 17th International Conference on Bioinformatics and Bioengineering, attempts to create a Convolutional Neural Network to classify Glioblastoma Multiforme Tumors based off of a BraTS 2017 dataset competition to produce a model that with the given factors of the dataset would predict the survival time of patients afflicted with this fatal cancer.
To start off with, Chato and Latifi took volumetric and location features form the dataset and calculated volumetric features of the tumor itself. Then they created a feature vector comprised of the volumetric data, the locations, and age were used to train several machine learning models, as seen in the Table 1 below.
They tried several classes for the survival prediction results such as short term, mid term, and long term survival rates, but ultimately decided to label short or long term for their models. Next, for their feature vector as seen from the table, they tested several models with 9, 7, and 4 features and data folds of either 5 or 15% removed to determine which model was best to continue their experiment. They deduced that with the two classes, their experiment would be conducted with a threshold of time of 18 months.
Next, they found the statistical and intensity texture features from the data set slice, with 155 slices in total to form the new feature vector. After some more model testing and a 10 fold accuracy, but found that their accuracies of the models would not exceed 46% even after trying to adjust the data with a PCA method.
Nextly, they extracted histogram features to train the data again and with KNN and SVM classifiers, they successfully raise the model accuracy to 65% between short and longterm survival classes.
Last but not least, they trained the AlexNet CNN after extracting deep features and using the MRI images from their dataset consisting of 3,703 samples and 913 samples for testing. They found with 5 fold accuracy, the Linear Discriminant model's accuracy was 91% and for an SVM classifier, 86.4%. Once the neural net was trained, they tested their model with a different dataset and found the model could not exceed 55% accuracy once again but found that the linear discriminant model continuously produced higher accuracy results of around 73%.