3293. Machine Learning to Differentiate Cerebral Tuberculosis from Brain Tumors Using Neuroimaging Features
Authors * Denotes Presenting Author
  1. Fatima Mubarak *; Aga Khan University
Cerebral tuberculosis (TB) often mimics primary and metastatic brain tumors, as well as other infectious pathologies of the brain, making it a challenging diagnosis to make. It is responsible for devastating sequelae and mortality, particularly in the developing world. There is a need for a rapid and accurate diagnostic approach to prevent the dismal outcomes arising as a result of delayed or incorrect diagnosis. We aim to develop a classifier to differentiate cerebral TB from brain tumors and other infections using various radiological features present on brain MRI with the help of machine learning.

Materials and Methods:
At our hospital, there were 72 cases of cerebral TB and 146 cases of non-TB (including meningiomas, gliomas, brain metastasis, fungal and bacterial brain infection) included and divided into training and test datasets. Features were selected using correlation matrix, and included radiological features recorded from brain MRI, such as ring enhancement, homogenous enhancement, basal meningeal enhancement, meningeal enhancement (not basal), homogeneous diffusion restriction, remote Infarcts, hydrocephalus, bilateral multi-focal lesions, unilateral multi-focal lesions, and multiple lesions within the same lobe, in addition to age and gender. After the application of Synthetic Minority Over-sampling Technique (SMOTE), SMOTE-Tomek Links, Edited Nearest Neighbor (ENN)SMOTE-ENN, and Adaptive Synthetic (ADASYN) techniques for balancing the datasets, classifier accuracy was tested using two models: logistic regression and random forest.

The highest accuracy (90.9%) was achieved using logistic regression along with SMOTE + TOMEK with 95.4% Area Under the Curve and an F1 score of 92.8%. Accuracy was increased by 6.81% after application of SMOTE + TOMEK to logistic regression models.

Machine learning shows promising role in clinical decision support systems for quickly and noninvasively differentiating brain tumors from infections. These classifiers can form the basis for mobile applications to be used in clinical settings. Sampling techniques should be employed to boost the performance of classifiers.