Discipline: Computer Sciences and Information Management
Subcategory: Biomedical Engineering
Session: 2
Room: Exhibit Hall A
Gabriella Tilahun - Alabama State University
Co-Author(s): Dr. Komal Vig, Alabama State University (ASU), Alabama
Epilepsy which is also known as seizure disorder is the occurrence of unpredictable seizure where regular nerve cell activity in the brain is disturbed. Brain injury, family tendency, or genetics can be mentioned as few causes although specific triggers are not yet clearly known. Patients can experience an episode to a cluster of seizures. A proper treatment varies on the severity and frequency of the seizure. Even with the help of today?s most advanced diagnosis, there is still room for error while detecting a region of the lesion in the brain. The gist of this project is to automate the process of identifying a specific part of the brain using deep learning algorithms, mainly the Random Forest algorithm to increase precision and accuracy. The approach we took included mainly three steps. The first step included parcellation of MRI scans of 56 patients over four atlases to generate row data using brain imaging software, Free Surfer. The row data had unnecessary features; thus, it had to be refined, which led to the Feature selection stage. A neurologist prepared a ground truth list after manually analyzing the data to identify features that had high magnitude variation among left and right brain patients. The data was then grouped into left and right brain patients. The mean and standard deviation of each feature was calculated to find the Fisher Separability constant. Features with higher separability constant were extracted from the raw data which were later used in the model training step. There was 65 ? 80% accuracy on different data sets. The last step was the model training. Some of the machine learning algorithms that came into account were SVM, KNN, and the Random Forest Algorithm. The Random Forest algorithm uses decision trees. In decision trees calculating nodes and forming rules takes place by Information gain and Gini-index calculation to find important features. The Random Forest is convenient for this classification because the problem of overfitting is eliminated since several samples are taken randomly to train more than one decision tree. A majority vote then made the predicted final result from the results of all the decision trees in the algorithm. The Random Forest is also found to be a good source for feature engineering. Out of 54 patients, 46 have useable MRI data, and there was 75- 80% classification accuracy on different data sets in the four atlases. The results obtained are promising for further studies. In conclusion, the project was successful in detecting the lateral region of the brain with a lesion. We are currently working on looking for better feature extraction method and modifying the algorithm.
Funder Acknowledgement(s): This work was supported by US Dept. of Education, The Minority Science and Engineering Improvement Program (MSEIP) (P120A150008) to Dr. Komal Vig (PD). And The Medical Imagine and Image Analysis Laboratory of Dr. Loew?s lab, The George Washington University (GWU).
Faculty Advisor: Rajendran Swamidurai, rswamidurai@alasu.edu
Role: My part in the project was writing source code in python for data extraction (feature engineering) and modifying then training the machine learning model.