Strike-Ms: Synergistic Texture and Raw Intensity Knowledge Extraction for Active Ms Lesion Detection on Non-Contrast Mri Publisher



Amini A ; Shayganfar A ; Amini Z ; Ostovar L ; Hajiahmadi S ; Chitsaz N ; Rabbani M ; Kafieh R
Source: IEEE Access Published:2025

Abstract

Multiple sclerosis (MS) is an autoimmune disease that affects the central nervous system. To detect active lesions in this disease, gadolinium-based contrast agents are injected before magnetic resonance imaging (MRI). These materials can have side effects for the patient after several injections. The purpose of this study is to use deep learning methods to detect active lesions in MRI images without using contrast agents. Our data included 9097 lesions from 130 patients from four different imaging centers. First, the lesions were identified, and each one was separated from the images as a patch. Texture features were also calculated for each lesion. Then, the study was conducted based on the input type in three different phases, including image intensity, texture features, and the combination of image intensity and texture features. In each phase, a proper CNN network was designed. In addition, 13 pre-trained CNN networks were also applied to predict lesions in the third phase. The ensemble technique was another method used in this phase to categorize lesions. Additionally, inter-center variability and cross-center generalizability assessment analyses were conducted to evaluate the model’s generalizability rigorously. The statistical results for the designed network of the first phase were obtained as an accuracy value of 85%, specificity of 95%, sensitivity of 74%, and AUC of 0.90. These values for the second phase were 77%, 90%, 63%, and 0.85, respectively, and for the third phase, 91%, 94%, 87%, and 0.94, respectively. Among the pre-trained networks in the third phase, the EfficientNetV2M showed the best performance with an average accuracy of 88%, specificity of 91%, sensitivity of 86%, and AUC of 0.92. This study indicates that while using image intensity and texture features separately effectively distinguishes active from inactive lesions, their combination yields even better results. © 2025 Elsevier B.V., All rights reserved.