Texture analysis and radiomics in the diagnosis of multiple sclerosis: a review

Full Text

INTRODUCTION

Multiple sclerosis (MS) is an autoimmune demyelinating and neurodegenerative disease of the central nervous system (CNS), that occurs as episodes of neurological deficit separated in time and is characterized by the formation of multiple lesions in the brain and spinal cord [1, 2].

Nearly 3 million people worldwide suffer from MS, which corresponds to 35.9 cases per 100,000. MS is one of the most common CNS diseases in young working-age individuals and ranks fourth among all neurological diseases. MS accounts for 36 to 78.5 cases per 100,000 in the Russian Federation. The disease can manifest at any age, but the median age of onset is 28.3–31.2 years, which gives it social significance and drives active interest among researchers [3, 4].

The exact etiology, pathogenesis, and mechanisms of disease progression remain incompletely understood [5]. It is believed that MS is a multifactorial process, the development of which is influenced by the interactions of immunological and genetic factors, as well as the environment [6, 7].

The diagnosis of MS is established based on a characteristic clinical picture. However, the diversity of manifestations often complicates early recognition of the disease, which leads to a delayed initiation of effective therapy and reducing patients' quality of life [8, 9].

The main additional diagnostic method for MS is magnetic resonance imaging (MRI), which allows detecting multifocal brain lesions and confirming the key characteristics: dissemination of the pathological process in space and time [1, 9].

Problems in the differential diagnosis of focal changes, as well as the frequent dissociation between clinical symptoms and MRI data, necessitate the search for more sensitive and specific damage biomarkers that can ensure early detection and prediction of the disease course [1, 10–15].

One of the promising additional methods for evaluating radiological images in the diagnosis of demyelinating diseases is the application of texture analysis and radiomics [16].

This review analyzes publications dedicated to the identification and implementation into clinical practice of MRI image biomarkers obtained using radiomics and texture analysis, which enable the differential diagnosis of demyelinating diseases, particularly MS.

Search Methodology

In this review, we conducted an analysis of scientific publications from 2016 to 2024 (the search depth was 9 years). The search was performed in the PubMed and eLibrary search engines using the keywords: "radiomics", "digital image texture analysis", "multiple sclerosis", and their Russian-language equivalents — "радиомика" (radiomics), "текстурный анализ" (texture analysis), and "рассеянный склероз" (multiple sclerosis).

This review includes only original studies dedicated to the application of radiomics and digital image texture analysis in the diagnosis of demyelinating diseases. Clinical and practical guidelines, monographs, abstracts, dissertations, papers unrelated to MS, and studies without the use of MRI were excluded.

Using the PubMed search engine, 17 original English-language studies were selected and analyzed (Table 1), whereas no Russian-language publications on the topic were found in eLibrary for the specified period.

 

Table 1. Publications dedicated to the application of radiomics and texture analysis in the diagnosis of demyelinating diseases

Authors	Year	Differential Diagnosis	Number of patients, n	MRI sequence	Software	Segmentation
Y. Liu et al. [24]	2019	Multiple sclerosis and neuromyelitis optica spectrum disorders (spinal cord)	189	T2-WI	MATLAB	Manual
G. Caruana et al. [37]	2020	Multiple sclerosis: acute active and chronic lesions	58	SWI, T2-WI	LIFEx	Manual
Y. Peng et al. [35]	2021	Multiple sclerosis: chronic active and inactive lesions	36	T2-FLAIR	PyRadiomics (3D Slicer)	Manual
C. E. Weber et al. [30]	2021	Multiple sclerosis: chronic active and inactive lesions	60	T1-MPRAGE	MaZda	Automatic
Z. Yan et al. [41]	2021	Multiple sclerosis and neuromyelitis optica spectrum disorders	83	QSM	PyRadiomics (Python)	Manual
T. He et al. [40]	2022	Multiple sclerosis and microangiopathy	952	T2-WI	IBEX	Manual
X. Luo et al. [42]	2022	Multiple sclerosis and systemic lupus erythematosus	112	T2-FLAIR	MATLAB	Manual
C. Fiscone et al. [54]	2023	White matter in multiple sclerosis and in healthy individuals	151	QSM	PyRadiomics (Python)	Automatic
B. Khajetash et al. [33]	2023	Multiple sclerosis: chronic active and inactive lesions	36	T2-FLAIR	PyRadiomics	Manual
T. Li et al. [39]	2023	Neuromyelitis optica spectrum disorders and anti-MOG- associated diseases	121	T2-WI, T2-FLAIR, T1-WI	PyRadiomics (Python)	Manual
H. Tavakoli et al. [34]	2023	Multiple sclerosis: acute active and chronic lesions	82	T2-FLAIR	PyRadiomics (3D Slicer)	Manual
R. Faustino et al. [32]	2024	Multiple sclerosis lesion, perifocal zone, and normally appearing matter (children)	11	T2-FLAIR	PyRadiomics (3D Slicer)	Semi-automatic
B. C. Kelly et al. [53]	2024	Multiple sclerosis: chronic active and inactive lesions	167	T2-FLAIR	PyRadiomics (3D Slicer)	Manual
X. Luo et al. [28]	2024	Multiple sclerosis, neuromyelitis optica spectrum disorders, and anti-MOG-associated diseases	195	T2-FLAIR	MATLAB	Manual
A. Rostami et al. [55]	2024	Multiple sclerosis: acute active and chronic lesions	134	T2-WI	PyRadiomics (3D Slicer)	Manual
F. Shekari et al. [36]	2024	Multiple sclerosis: acute active and chronic lesions	34	DWI	Python	Manual
Z. Shi et al. [31]	2024	Multiple sclerosis: acute active lesions, chronic active and inactive lesions	45	T2-FLAIR	PyRadiomics (Python)	Manual

Note. MRI, magnetic resonance imaging; T2-WI, T2-weighted imaging; SWI, Susceptibility Weighted Imaging; T2-FLAIR, T2-Weighted Fluid-Attenuated Inversion Recovery; T1-MPRAGE, T1-Magnetization Prepared Rapid Acquisition Gradient Echo; QSM, Quantitative Susceptibility Mapping; T1-WI, T1-weighted imaging; DWI, diffusion-weighted imaging; anti-MOG-associated diseases, myelin oligodendrocyte glycoprotein antibody-associated diseases.

 

BASIC CONCEPTS IN RADIOMICS

Biomedical images obtained using various methods typically contain from 4096 to 65536 shades of gray. However, the human eye under optimal conditions is capable of perceiving only 1000 of them. Furthermore, image analysis is often limited by the use of displays (monitors) with a resolution of 64 or 256 shades of gray, as well as the capabilities of visualization applications [17].

Since the early 1980s, methods for extracting relevant information from various medical images have been actively investigated. One such method is the application of texture analysis [18–20].

It represents a method of objective assessment of medical images, allowing to calculate parameters of pixel and voxel value distribution and their relationships, which cannot be detected during visual observation [16, 21].

The concept of radiomics was formulated later, partly due to the work of the Dutch scientist Lambin, in 2012 [22, 23].

Radiomics is the generalized process of automatic extraction, test, and subsequent interpretation of quantitative features (image biomarkers) based on the specific characteristics of the analyzed texture data [22, 24].

Texture analysis and radiomics represent a hybrid analytical process at the intersection of mathematical, informational, statistical, and medical sciences [25]. The process of applying radiomic analysis can be divided into several stages (Fig. 1).

 

Fig. 1. Stages of texture and radiomic analysis (schematic representation proposed by the authors).

 

Image acquisition. The results of radiomic analysis depend on the source data, which are determined by the biomedical imaging techniques. This can be radiography, ultrasound, computed tomography, MRI, or nuclear medicine techniques. That is why data can vary both between different methods and within a single one, depending on the protocols and the equipment used [26].

Identification of changes of interest. This stage aims to detect and classify changes that may have diagnostic and prognostic value and reflect the specifics of a particular pathological process [26].

Segmentation (outlining the boundaries of the region of interest using computer contouring). At this stage, the boundaries of the area that will be tested and included in the region of interest (ROI) are outlined. The determination of the ROI can be performed using manual, semi-automatic, and automatic methods using three-dimensional (3D-ROI) or two-dimensional (2D-ROI) image reconstructions [16, 21, 27].

Obtaining texture features. The extraction of quantitative texture features is carried out automatically using software. Such well-known tools include: PyRadiomics® (Computational Imaging & Bioinformatics Lab, USA), Medical Imaging Interaction Toolkit® (MITK) (German Cancer Research Center, Germany), LIFEx® (Institut Curie, France), Standardized Environment for Radiomics Analysis® (SERA) (Johns Hopkins University, USA), Cancer Imaging Phenomics Toolkit® (CaPTk) (Center for Biomedical Image Computing and Analytics, USA), Imaging Biomarker Explorer® (IBEX) (MD Anderson Cancer Center, USA), etc. [28, 29].

Statistical test of the obtained texture features and building of diagnostic models. All texture features are subdivided into the following groups [15, 27]:

• first-order texture features, namely histogram characteristics reflecting information related to the distribution of gray level (shades) within the ROI without taking into account the spatial relationships between pixels. These parameters include: maximum, minimum, mean, and median intensity values in the selected region, root mean square deviation, distribution asymmetry, kurtosis, entropy, skewness, etc.;
• second-order texture features, which take into account the spatial relationships of pixels and voxels, and reflect their distribution both individually and in groups in all directions of three-dimensional space. These features are derived from radiomic matrices compiled based on specific ranges of image intensities. Such matrices include: Gray Level Run Length Matrix (GLRLM); Gray Level Size Zone Matrix (GLZLM); Gray Level Co-occurrence Matrix (GLCM); Neighboring Gray Tone Difference Matrix (NGLDM), etc.

Based on first- and second-order data, higher-level features are formed using mathematical operations such as the Fourier transform, wavelet analysis, and various filters. These operations allow describing the statistical characteristics of the images.

In addition, for 3D analysis, shape characteristics of a given ROI are also calculated, describing its area, volume, maximum linear dimension, compactness, sphericity, and the ratios between these characteristics and other parameters [16, 29].

From the multitude of obtained radiomic data, using the built mathematical and statistical models, the most informative and statistically significant texture features — image biomarkers — are selected. They can be used alone or in combination with the patient's demographic, histological, and genomic data to establish a clinical diagnosis and select personalized therapy [16].

APPLICATION OF RADIOMICS AND TEXTURE ANALYSIS IN MULTIPLE SCLEROSIS

Among the studies dedicated to texture analysis and radiomics in the MS diagnosis is the work by C. E. Weber et al. [30]. The authors proposed a diagnostic model based on the test of T1-MPRAGE MRI images¹. The aim of this model was the differentiation of chronic active, or expanding ("smoldering"), and inactive (stable) demyelinating lesions. The study identified and analyzed 36 radiomic features extracted for each of the 370 lesions detected in 60 patients. Of all the texture parameters, four were the most significant. As a result of the statistical analysis, a diagnostic model was created that was able to distinguish between the two types of lesions with good diagnostic characteristics: area under the curve (AUC) — 0.7.

Many authors emphasized the significance of the T2-FLAIR² mode for extracting image texture biomarkers when evaluating focal changes in the brain. In particular, Z. Shi et al. [31] proposed distinguishing three types of demyelination lesions in MS: acute active (contrast-enhancing), chronic active, and inactive lesions. The authors analyzed 140 radiomic features obtained for each of the 614 lesions detected in 45 patients. As a result, a model was created that distinguished between chronic active and inactive lesions with very good diagnostic accuracy (AUC = 0.893). In turn, the model capable of differentiating all three types of lesions possessed excellent diagnostic characteristics (AUC = 0.920).

Furthermore, the importance of using the T2-FLAIR² sequence in assessing demyelinating changes in the brain is confirmed by other studies [32–35]. Thus, R. Faustino et al. [32] identified the ten most significant biomarkers with AUC > 85%, which allow distinguishing «border zones» (tissue around lesions up to 1 mm thick), areas of normally appearing white matter, and brain lesions in children with MS. Other studies have proposed signatures that allow identifying progressive (expanding) and acute active lesions in MS [32, 34]. Thus, B. Khajetash et al. [33] proposed a technique that showed excellent diagnostic characteristics: AUC — 0.85, sensitivity — 0.85, and specificity — 0.66. In turn, the methodology by H. Tavakoli et al. [34] demonstrated excellent results: AUC — 0.92, sensitivity — 0.86 and specificity — 0.84. Additionally, in the study Y. Peng et al. [35] MS lesion progression was determined using a machine learning model based on radiomic data. It showed really good effectiveness, reaching an AUC value of 0.857.

The efficacy of using diffusion-weighted imaging in the differential diagnosis of acute active and chronic (non-contrast-enhancing) MS lesions was demonstrated by F. Shekari et al. [36]. The authors extracted and analyzed 89 radiomic features for each of the 255 lesions found in 34 patients. The created diagnostic models, incorporating from 13 to 20 of the most relevant biomarkers, differentiated acute and chronic MS lesions with excellent diagnostic characteristics (AUC = 0.957).

G. Caruana et al. [37] performed texture analysis of Susceptibility Weighted Imaging (SWI) and T2-weighted images (T2-WI), to identify acute active and chronic lesions in MS. The authors analyzed 35 radiomic features obtained for each of the 175 lesions identified in 58 patients. The model created based on SWI images was able to differentiate acute and chronic lesions with good diagnostic characteristics (AUC = 0.778), unlike the model built on T2-WI, where the authors did not obtain notable differences between the two types of lesions (AUC = 0.567). However, N. Michoux et al. [38] in an earlier work demonstrated the effectiveness of a diagnostic model based on T2-WI for the classification of acute active and chronic MS lesions. The authors analyzed 21 radiomic features obtained for each of the 81 lesions identified in 21 patients. The model evaluating the diagnostic capabilities of individual imaging biomarkers showed results ranging from average (AUC = 0.638) to very good (AUC = 0.835). In a combined model, where T. Li et al [39] used sequences such as T2-WI, T2-FLAIR², and T1-WI for the differential diagnosis of myelin oligodendrocyte glycoprotein antibody-associated diseases (anti-MOG-associated diseases) in children, T2-WI played a key role. It showed the best diagnostic performance when used separately (AUC = 0.83) and also revealed the highest number of biomarkers (n = 22). The combined model, incorporating 32 biomarkers out of 1688 radiomic features, possessed excellent diagnostic characteristics (AUC = 0.905).

In addition, T. He et al. [40] also demonstrated the effecacy of using T2-WI to obtain texture parameters for the differential diagnosis of MS lesions and microangiopathy. The authors identified 9 of the 472 most significant biomarkers obtained for each of the 3068 lesions detected in 952 patients. The final model capable of distinguishing between the two types of lesions, possessed excellent diagnostic characteristics (AUC = 0.9).

The significance of using Quantitative Susceptibility Mapping (QSM) technique to extract texture biomarkers with their subsequent use in the differential diagnosis of MS and neuromyelitis optica spectrum disorders (NMOSD) was demonstrated by Z. Yan et al. [41]. The researchers analyzed seven ROI in the subcortical structures. As a result, they obtained 874 radiomic features for each of the 7 zones in 83 patients. The diagnostic model, incorporating two to four of the most relevant biomarkers showed results ranging from good (AUC = 0.702) to excellent (AUC = 0.902).

Differential diagnosis of focal changes in the group of demyelinating diseases in MS, NMOSD, and anti-MOG-associated diseases was conducted by X. Luo et al. [28]. The authors analyzed 378 radiomic features obtained for each of the 212 ROI in 195 patients. The diagnostic model based on T2-FLAIR² MRI images differentiated MS lesions using the most significant biomarkers with excellent diagnostic characteristics (AUC = 0.909).

The possibility of applying radiomics in the differential diagnosis of focal changes in the brain matter in MS and systemic lupus erythematosus was demonstrated in the work of X. Luo et al. [42], wherein the authors identified and analyzed 371 radiomic biomarkers obtained for focal changes in 112 patients in T2-FLAIR² mode. They used a machine learning model based on radiomic data, which showed excellent efficacy, achieving an AUC value of 0.967.

A test of current research (2016–2024) demonstrates a growing interest in the application of radiomics in the diagnosis of MS and related diseases. The most frequently used MRI mode was T2-FLAIR² (53% of studies), while T2-WI and QSM were used less frequently (29% and 12%, respectively). Segmentation of ROI was predominantly performed manually (82%), less frequently — automatically (12%) or semi-automatically (6%). Study sample sizes ranged from 11 to 952 patients. Specialized software extracted from 21 to 874 radiomic features, from which different statistical models identified as few as four of the most significant image biomarkers. The most significant were second-order biomarkers, including:

• GLCM accounts for the arrangement of pixel pairs in four directions to calculate texture indices (e.g., the arithmetic mean of the number of pixel pairs, the frequency of occurrence of identical values, the degree of linear dependence between pixel pairs, etc.);
• GLRLM is created by placing different gray intensities on a single matrix axis and reflects the frequency of occurrence of elements with the same gray intensity in a row;
• GLZLM describes the size of homogeneous zones for each gray level. Unlike GLRLM, this biomarker calculates the number of adjacent elements with the same gray intensity in three dimensions;
• NGLDM describes the differences between each image element and its nearest "neighbors" (8 elements around a pixel in a 2D image or 26 voxels in a 3D image).

LIMITATIONS OF TEXTURE ANALYSIS AND RADIOMICS

When using texture analysis and radiomics methods to evaluate MRI images, there are certain technical difficulties and limitations. These include the diversity of approaches to radiomic analysis at each of its stages, which complicates the comparison of results from different studies [43].

During image acquisition, to ensure the consistency of results and their clinical applicability under ideal conditions, it is necessary to adhere to a unified scanning protocol, the regulation of which can reduce the variance in image quality and technical parameters that can significantly affect the derived metrics. Furthermore, MRI uses arbitrary units to measure MR signal intensity [44, 45]. This fact underscores the need for study preprocessing, which includes the processes of normalization, standardization, and discretization of MRI images [46, 47].

The most important, complex, and crucial stage in radiomics is the segmentation process. The difficulty in contouring MS lesions is due to their diverse shapes, sizes, MR signal intensity, and blurred contours due to edema during the active stage of the inflammatory process [48]. In addition, the interpretation and segmentation of each study by several specialists, along with conducting a reference test, is important, as it significantly increases the diagnostic accuracy of the obtained metrics [44, 49].

At the stage of extracting texture features, various radiomic analysis software typically generates a large and unequal number of features, which reduces their reproducibility and limits the possibilities for generalized application. To solve this issue, an independent international collaboration was created — the Image Biomarker Standardisation Initiative (IBSI), aimed at standardizing the calculation of radiomic characteristics using reference phantoms and stated values, as well as evaluating the software platforms used for radiomic analysis [50].

To assess the effectiveness of the entire radiomics workflow, a special quality metric was developed and implemented — the RQS (Radiomics Quality Score). It represents a total percentage score assigned for the correct execution of each of the study stages [46, 51].

During statistical test, various methods for building diagnostic models are applied, which creates difficulties in harmonizing and comparing study results. In classical works, data are presented in the form of tables, which are then analyzed using traditional statistical models. Furthermore, radiomics actively employs machine learning models and neural networks, which offer a multitude of methods for testing and evaluating both quantitative and qualitative data and their interrelationships. Additionally, some works use deep learning models that represent a comprehensive approach, combining the entire image processing workflow into a single model, predicting results directly from raw images without human intervention [46, 52].

CONCLUSION

Neuroimaging is the most promising area for the application of texture analysis methods and radiomics. Despite existing difficulties and mixed results, these approaches represent a promising tool for the supplementary evaluation of multifocal brain changes in demyelinating diseases. Creating an optimal algorithm for calculating texture features, identifying the most relevant ones, and standardization and validation the derived biomarkers can ensure the integration of these methods into clinical practice and improve the diagnostic accuracy of demyelinating changes, particularly in MS.

ADDITIONAL INFORMATION

Author contributions: G.I. Khvastochenko: data curation, writing — original draft, writing — review & editing, visualization; V.V. Bryukhov, M.V. Krotenkova: data curation, writing — original draft. All the authors approved the version of the manuscript to be published and agreed to be accountable for all aspects of the work, ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Ethics approval: Not applicable.

Funding sources: No funding.

Disclosure of interests: The authors have no relationships, activities, or interests for the last three years related to for-profit or not-for-profit third parties whose interests may be affected by the content of the article.

Statement of originality: No previously published material (text, images, or data) was used in this study or article.

Data availability statement: The editorial policy regarding data sharing does not apply to this work.

Generative AI: No generative artificial intelligence technologies were used to prepare this article.

Provenance and peer-review: This article was submitted unsolicited and reviewed following the standard procedure. The peer-review process involved one external reviewer and two members of the Editorial Board.

¹ T1-MPRAGE (T1-Magnetization Prepared Rapid Acquisition Gradient Echo) is a three-dimensional pulse sequence with magnetization preparation and fast gradient echo.

² T2-FLAIR (T2-Weighted Fluid-Attenuated Inversion Recovery) is a T2-weighted image with fluid signal suppression using inversion.

About the authors

Gleb I. Khvastochenko

Russian Center of Neurology and Neurosciences

Author for correspondence.
Email: hvastochenko.g.i@neurology.ru
ORCID iD: 0009-0003-4628-3069
SPIN-code: 8988-6959
Russian Federation, Moscow

Vasiliy V. Bryukhov

Russian Center of Neurology and Neurosciences

Email: abdomen@rambler.ru
ORCID iD: 0000-0002-1645-6526
SPIN-code: 6299-3604

MD, Cand. Sci. (Medicine)

Russian Federation, Moscow

Marina V. Krotenkova

Russian Center of Neurology and Neurosciences

Email: krotenkova_mrt@mail.ru
ORCID iD: 0000-0003-3820-4554
SPIN-code: 9663-8828

MD, Dr. Sci. (Medicine)

Russian Federation, Moscow

References

Supplementary files