نهمین کنگره نقشه برداری مغز

نمایش مقاله

Title:

A Deep Learning Approach for Discriminating Active Multiple Sclerosis Lesions Using Texture Features in Non-Contrast MRI

Subject:

Artificial Intelligence and Brain Mapping

Presenter Author:
Name:	AmirAbbas Amini

Author:
Name 1:	AmirAbbas Amini - Corresponding - Presenter
Affiliation 1:	School of Advanced Technologies in Medicine, Medical Image and Signal Processing Research Center, Isfahan University of Medical Sciences, Isfahan, Iran

Name 2:	Rahele Kafieh
Affiliation 2:	Department of Engineering, Durham University, Durham, UK

Name 3:	Azin Shayganfar
Affiliation 3:	Department of Radiology, School of Medicine, Isfahan University of Medical Sciences, Isfahan, Iran

Name 4:	Zahra Amini
Affiliation 4:	School of Advanced Technologies in Medicine, Medical Image and Signal Processing Research Center, Isfahan University of Medical Sciences, Isfahan, Iran

Name 5:	Leila Ostovar
Affiliation 5:	Department of Radiology, School of Medicine, Isfahan University of Medical Sciences, Isfahan, Iran

Name 6:	Somayeh Haji Ahmadi
Affiliation 6:	Department of Radiology, School of Medicine, Isfahan University of Medical Sciences, Isfahan, Iran

Name 7:
Affiliation 7:

Name 8:
Affiliation 8:

Name 9:
Affiliation 9:

Name 10:
Affiliation 10:

Abstract

Original Research

introduction :

Multiple sclerosis (MS) is a chronic demyelinating disorder of the central nervous system. Accurately distinguishing between active and inactive lesions is essential for monitoring disease progression and guiding treatment. Current clinical practice often relies on gadolinium-enhanced MRI, which carries potential risks and limitations. This study explores a non-contrast alternative by applying texture feature analysis from FLAIR MRI sequences integrated into a deep convolutional neural network (CNN) for automated lesion classification.

Methods:

The study included a dataset of 9097 MRI images from 130 MS patients across four imaging centers in Isfahan, Iran. All lesions were identified, masked, and labeled before extraction as image patches. Three groups of image sets were prepared: I) lesions without mask dilation, II) lesions with masks dilated by 3 pixels, and III) lesions with masks dilated by 6 pixels. Texture features including contrast, dissimilarity, homogeneity, angular second moment (ASM), energy, and correlation were extracted using the Gray Level Co-occurrence Matrix (GLCM) method at distances of 0 and 1 and angles of 0 and 90 degrees. A custom deep learning CNN was then designed and trained to classify lesions as active or inactive. The pipeline was evaluated using 5-fold cross-validation across all three groups.

Results:

The CNN model demonstrated varying performance across the three tested conditions. For lesions without mask dilation, the model achieved an accuracy of 72% (standard deviation ± 0.38), sensitivity of 65% (± 2.42), specificity of 80% (± 2.65), and an AUC of 0.81 (± 0.02). With a dilation size of 3 pixels, performance improved significantly, yielding an accuracy of 80% (± 0.74), sensitivity of 70% (± 2.17), specificity of 90% (± 2.17), and an AUC of 0.87 (± 0.03). Using a dilation size of 6 pixels also showed strong results, with an accuracy of 78% (± 0.82), sensitivity of 69% (± 2.31), specificity of 86% (± 2.34), and an AUC of 0.85 (± 0.02).

Conclusion:

The findings indicate that texture features derived from GLCM can effectively support deep learning models in classifying MS lesions without gadolinium contrast. The highest performance was achieved using a dilation size of 3 pixels, underscoring the diagnostic value of incorporating tissue immediately surrounding the lesion. This non-contrast, deep learning-based approach provides a promising, automated tool for identifying active MS lesions in clinical practice.

Keywords:

Multiple Sclerosis(MS), Deep Learning, Texture Features, Non-Contrast MRI, Lesion Classification