Classification of Active Multiple Sclerosis Lesions in MRI Without the Aid of Gadolinium-Based Contrast Using Textural and Enhanced Features from FLAIR Images

Author(s):Show less -

Freire, Paulo G. L. ; Idagawa, Marcos Hideki ; Lobato de Oliveira, Enedina Maria ; Abdala, Nitamar ; Carrete Jr, Henrique ; Ferrari, Ricardo J. ; Gervasi, O ; Murgante, B ; Misra, S ; Garau, C ; Blecic, I ; Taniar, D ; Apduhan, BO ; Rocha, AMAC ; Tarantino, E ; Torre, CM ; Karaca, Y

• Document type: Journal article
• Source: COMPUTATIONAL SCIENCE AND ITS APPLICATIONS - ICCSA 2020, PT II; v. 12250, p. 15-pg., 2020-01-01.

Abstract

Multiple sclerosis (MS) is an autoimmune demyelinating disease that affects one's central nervous system. The disease has a number lesion states. One of them is known as active, or enhancing, and indicates that a lesion is under an inflammatory condition. This specific case is of interest to radiologists since it is commonly associated with the period of time a patient suffers most from the effects of MS. To identify which lesions are active, a Gadolinium-based contrast is injected in the patient prior to a magnetic resonance imaging procedure. The properties of the contrast medium allow it to enhance active lesions, making them distinguishable from nonactive ones in T1-w images. However, studies from various research groups in recent years indicate that Gadolinium-based contrasts tend to accumulate in the body after a number of injections. Since a comprehensive understanding of this accumulation is not yet available, medical agencies around the world have been restricting its usage to cases only where it is absolutely necessary. In this work we propose a supervised algorithm to distinguish active from nonactive lesions in FLAIR images, thus eliminating the need for contrast injections altogether. The classification task was performed using textural and enhanced features as input to the XGBoost classifier on a voxel level. Our database comprised 54 MS patients (33 with active lesions and 21 with nonactive ones) with a total of 22 textural and enhanced features obtained from Run Length and Gray Level Co-occurrence Matrices. The average precision, recall and F1-score results in a 6-fold crossvalidation for active and nonactive classes were 0.892, 0.968, 0.924 and 0.994, 0.987, 0.991, respectively. Moreover, from a lesion perspective, the algorithm misclassified only 3 active lesions out of 157. These results indicate our tool can be used by physicians to get information about active MS lesions in FLAIR images without using any kind of contrast, thus improving one's health and also reducing the cost of MRI procedures for MS patients. (AU)

• FAPESP's process: 18/08826-9 - Development of feature engineering and deep learning techniques applied to the classification of magnetic resonance images in healthy cognitive aging, mild cognitive impairment and Alzheimer's Disease
• Grantee: Ricardo José Ferrari
• Support Opportunities: Regular Research Grants
• FAPESP's process: 16/15661-0 - Automatic computational scheme for segmentation and identification of acute inflammatory process in multiple sclerosis lesions without using contrast agent
• Grantee: Paulo Guilherme de Lima Freire
• Support Opportunities: Scholarships in Brazil - Doctorate