Evaluation of Classifiers for the Identification of Multiple Sclerosis Lesions in Neural MRI Scans with Attributes Extracted from Pre-trained Neural Networks

About 2.8 million people worldwide have multiple sclerosis, and around 250,000 new cases are diagnosed annually. This disease damages the myelin sheath of neurons, injuring electrical signal conduction, and causing impairment and loss of senses, movement, and other neurological functions. Although the disease has no cure, early diagnosis is essential for the initiation of adequate treatment, controlling eventual outbreaks, delaying the advance, and improving the quality of life of patients. Neuroaxis magnetic resonance imaging is used to investigate this disease, identifying and following up lesions in the brain and spinal cord tissues; however, the diagnosis only through the visual evaluation of these exams may lack information for a quick and accurate analysis. Recent studies present computational methods based on artificial intelligence that allow the identification of lesions caused by the disease, aiming to overcome these visual limitations, but still have limitations in terms of accuracy and scope. The objective of this work was to evaluate the potential of classifiers based on machine learning algorithms in the identification of multiple sclerosis lesions in brain tissue. Axial FLAIR MRI brain exams were used from two databases combining filtering, normalization, and enhancement image pre-processing methods to extract the exam attributes. The algorithms evaluated were Decision Tree, Random Forest, K-Nearest Neighbors, Support Vector Machine and Logistic Regression, through training with attributes extracted from 3 pre-trained neural models: SqueezeNet, Inception V3 and VGG-19. The classifiers were evaluated by the classification of images from the studies as “with lesion” and “without lesion”, and the SVM classifier trained with attributes extracted from the pre-trained neural model Inception V3 provided the best result, obtaining AUC = 0.988, accuracy = 0.980, sensitivity = 0.990 and F1-Score = 0.979.