Association of Spinal Cord Radiomic Features and Disability in Multiple Sclerosis

ABSTRACT

Background and Purpose

Spinal cord pathology underpins disability accumulation in people with multiple sclerosis (pwMS). Visual inspection of spinal cord magentic resonance imaging (MRI) often fails to reliably detect injury. Radiomics analyzes signal intensities in images to identify pathological changes that may be imperceptible to the human eye. This study evaluated the application of radiomics to spinal cord MRI to distinguish subgroups of pwMS and disability correlations.

Methods

Radiomic features were extracted from upper cervical cord coverage on cross‐sectional 3.0T brain noncontrasted T1‐weighted MRI scans in pwMS and healthy controls (HCs). Ninety‐three radiomic features—predominantly gray‐level matrices—were extracted using Pyradiomics, with pixel heterogeneity considered to reflect neuroaxonal pathology. T2 lesion and brain substructure volumes were segmented from 3D fluid‐attenuated inversion recovery and magnetization‐prepared rapid gradient‐echo sequences using an in‐house 2.5D U‐Net convolutional neural network to encapsulate neuroinflammation and neurodegeneration. Cervical cross‐sectional area (C1−C3) was measured using in‐house atlas‐based segmentation. Imaging features were compared between pwMS and HCs, and pwMS by phenotype (relapsing vs. progressive), age, and race. Associations of imaging features with Patient‐Determined Disease Steps (PDDS) were examined.

Results

Among 2966 pwMS and 41 HCs, we identified radiomic features distinguishing pwMS from HCs, and pwMS by phenotype, age, and race. Radiomic features exhibited stronger correlations with PDDS than conventional MRI measures.

Conclusions

Radiomics identified pathological changes in pwMS in varying stages of the disease course that are undetectable by conventional spinal cord MRI. Radiomics may increase the yield of spinal cord MRI in pwMS and serve as biomarkers predicting disability worsening.

1. Introduction

Multiple sclerosis (MS) is an immune‐mediated disorder of the central nervous system  (CNS), in which neuroinflammation and neurodegeneration drive disability accumulation [1, 2]. Magnetic resonance imaging (MRI) has a well‐established role in MS. Advances in brain MRI have facilitated visualization and recapitulation of neuroinflammatory pathology in people with multiple sclerosis (pwMS) [3, 4]. Despite this, there remains a large discrepancy between the degree of disease activity on MRI and clinical disability [5, 6]. This clinical radiologic paradox suggests a potential limit on the amount of data extractable through conventional MRI that meaningfully contributes to explaining global disability.

Spinal cord pathology is very common in MS and appears to be highly relevant to disability [7, 8]. However, visual inspection of conventional spinal cord MRI lacks sensitivity and reliability, largely related to technical issues given the small caliber of the cord and associated physiological mobility [9, 10]. There is, therefore, an unmet need for new technologies to better image this structurally eloquent area.

Radiomics is increasingly utilized to analyze medical images. The radiomics approach extracts features from images based on signal intensities and inter‐relationships between pixels [11, 12]. The detected heterogeneity within regions of interest (ROI) is thought to reflect physiological and pathological changes in tissues, many of which may be undetectable by traditional visual inspection. The features extracted can be predefined based on clinical significance. Several studies have demonstrated the ability of radiomics to distinguish between MS and other demyelinating disorders [13, 14, 15, 16, 17]. However, there is a paucity of data evaluating radiomic features among subgroups of pwMS that are known to exhibit distinct pathology—for example, pwMS with a relapsing versus progressive course, younger versus older pwMS, and African American (AA) versus non‐AA pwMS [18, 19, 20]. Additionally, while radiomic features extracted from brain MRI have been shown to correlate with MS‐related clinical disability, the association of spinal cord radiomic features with clinical outcomes among pwMS has not yet been evaluated [21, 22].

The objective of this study was to evaluate spinal cord radiomics in MS. We aimed to identify radiomic features that distinguish pwMS from healthy controls (HCs), as well as differentiating between subgroups of pwMS by age, race, and relapsing versus progressive course. We then aimed to assess the association of radiomic features with objective neuroperformance and patient‐reported disability outcomes in comparison to conventional MRI measures.

2. Methods

2.1. Inclusion Criteria and Study Design

This was a retrospective, cross‐sectional cohort study of 2966 pwMS aged ≥16 years who underwent brain MRI and clinical assessment performed within 3 months of MRI across all Cleveland Clinic centers between December 2015 and February 2022. The diagnosis of MS was confirmed by the treating neurologist, according to the 2017 revised McDonald criteria, and the clinical course was categorized as relapsing or progressive MS [23, 24]. Forty‐one age‐ and sex‐matched HCs with brain MRI studies on the same MRI protocol as pwMS were used for comparison. Individuals with comorbid neurologic or severe systemic disorders were excluded. The study was conducted in accordance with the Declaration of Helsinki. Ethical approval was granted by the local Institutional Review Board (18‐1542).

2.2. MRI Acquisition and Analysis

Brain MRI scans were performed using a 3.0T scanner (Siemens, Erlangen, Germany), including coverage of the upper cervical spinal cord (C1−C3). Scans used standardized acquisition, including 3D Fluid Attenuated Inversion Recovery (FLAIR; echo time [TE] = 393 ms, repetition time [TR] = 5000 ms, inversion time [TI] = 1800 ms, flip angle [FA] = 120⁰, 1‐mm isotropic voxel size) and T1‐weighted magnetization prepared rapid gradient echo (MPRAGE; TE = 3 ms, TR = 7.8 ms; TI = 1000 ms, FA = 12°, 1‐mm isotropic voxel size) sequences. Only noncontrasted scans were analyzed. Preprocessing included iterative application of an N3 nonuniformity correction algorithm, intra‐time point registration using normalized mutual information (minctracc, Medical Image Numerical Computing [MINC]), and intensity normalization using the MINC toolkit, which used a histogram matching [25, 26]. The MPRAGE intensities were normalized with respect to the normal‐appearing white matter using the volume_pol tool (MINC). Using an in‐house 2.5D U‐Net convolutional neural network, T2 lesions and brain tissues were fully automatically segmented from the T1‐weighted and FLAIR images [27]. Briefly, the reference segmentations were generated by a single expert (KN), who semi‐automatically segmented T2 lesions on FLAIR for the ground truth segmentation in the convolutional neural network model. The semi‐automatic process involved a brush tool that segmented hyperintense objects. Brain structures (thalamus, hippocampus, caudate, pallidum, cerebral gray matter, cerebral white matter, cerebellar gray matter, cerebellar white matter, and lateral ventricles) were segmented using FreeSurfer 6.1 and manually checked for accuracy for reference segmentation in the model training [28]. The model was created using a training data set containing 149 subjects, a validation data set containing 20 subjects, and a testing data set containing 10 subjects. The Dice coefficient was used to measure the pixel‐wise similarity of predicted segmentation. For lesions, the mean Dice was 0.72 in validation and 0.69 in the testing data set. Analysis of lesion volumes on the same model demonstrated that reference lesion volume and predicted lesion volumes were highly correlated (r = 0.99; p < 0.001). Dice was 0.91 for the cerebral cortex, 0.95 for normal‐appearing white matter, 0.90 for cerebellar white matter, and 0.94 for lateral ventricles. Outer contour volumes were determined from the brain parenchyma (gray matter and white matter), and fractional volumes were calculated for each tissue and structure (e.g., whole brain fraction and gray matter fraction). The cervical cross‐sectional area was measured using an in‐house atlas‐based segmentation method for the ROI (C1−C3) [29, 30, 31].

2.3. Radiomic Feature Extraction

Ninety‐three radiomic features were extracted from a preidentified C1−C3 ROI of MPRAGE images through an open‐source package (PyRadiomics, Version 3.0; Boston, MA, USA; https://pyradiomics.readthedocs.io/en/latest/; Figure 1) [16, 32]. All features from classes of first‐order (histogram‐based) and higher‐order statistics (gray‐level co‐occurrence matrix [GLCM], gray‐level run length matrix [GLRLM], gray‐level size zone matrix [GLSZM], neighboring gray tone difference matrix [NGTDM], and gray‐level dependence matrix [GLDM]) were included in analyses [33, 34, 35, 36, 37]. Shape features were excluded from analyses due to their comparatively low robustness and reproducibility [38, 39].

FIGURE 1.

Examples of radiomic feature extraction from three sample patients (A, B, and C). Radiomic features were extracted from a preidentified region of interest of T1‐weighted MPRAGE images through an open‐source package (PyRadiomics). A1, B1, and C1 are axial views; A2, B2, and C2 are sagittal views. Percentiles refer to the feature value for each patient in relation to the entire cohort. GLCM, gray‐level co‐occurrence matrix; GLDM, gray‐level dependence matrix; GLRLM, gray‐level run‐length matrix; GLSZM, gray‐level size‐zone matrix; MPRAGE, magnetization prepared rapid gradient echo.

2.4. Clinical Outcomes

Disability was assessed using the Patient‐Determined Disease Steps (PDDS) and Multiple Sclerosis Performance Test (MSPT). The PDDS is a patient‐reported measure of global disability, with nine levels ranging from 0 (normal) to 8 (bedridden) [40, 41]. Due to small numbers of participants reporting certain scores, pwMS were categorized into quartiles of PDDS scores. The MSPT is an iPad‐based evolution of the Multiple Sclerosis Functional Composite, comprising the Walking Speed Test (WST), Manual Dexterity Test (MDT), Processing Speed Test (PST), and Contrast Sensitivity Test (CST), with z‐scores calculated based on normative data [42].

2.5. Statistical Analyses

Linear regression models were conducted to determine if radiomic and conventional MRI features were different between pwMS and HCs, and among pwMS between age quartiles (highest vs. lowest quartile), race (AA vs. non‐AA), and MS course (relapsing vs. progressive). A linear regression model was chosen to allow adjustment for age, sex, race, disease duration, MS course, and body mass index. Multivariable linear and logistic regression models were then fitted with radiomic and conventional MRI features as predictors, and MSPT z‐scores and PDDS (highest vs. lowest quartile) as dependent variables. Spearman correlations between disability outcomes and radiomic and conventional MRI features were examined. p‐Values were calculated using a Holm−Bonferroni correction to account for multiple comparisons. Statistical analyses were performed in R software, version 4.4.2 (Vienna, Austria; https://www.R‐project.org/) [43].

3. Results

3.1. Baseline Demographic and Clinical Characteristics

A total of 2966 pwMS were included. Mean age was 47.0 years (standard deviation [SD] = 12.8), 73.3% were female, 82.5% were Caucasian, 72.6% had relapsing disease, and the median disease duration was 11.4 years (interquartile range = 4.8, 19.7; Table 1). Thirty‐eight percent of pwMS had no disability, as measured by PDDS. Forty‐one HCs were included, with a mean age (43.8 years [SD = 9.3]) and sex (75.6% female) similar to the MS cohort.

TABLE 1.

Baseline demographic and clinical characteristics among people with multiple sclerosis.

Participants, n	2966
Age, years, n (%)
first quartile (16.8−38.3)	740 (24.9)
second quartile (38.3–47.1)	739 (24.9)
third quartile (47.1–55.9)	742 (25.0)
fourth quartile (55.9–85.1)	745 (25.1)
Sex, n (%)
Female	2173 (73.3)
Race, n (%)
Caucasian	2448 (82.5)
African American	331 (11.1)
Other	187 (6.3)
Body mass index, mean (SD)	29.4 (7.4)
MS course, n (%)
Relapsing	2152 (72.6)
Progressive	814 (27.4)
Disease duration, years, median (IQR)	11.4 (4.8, 19.7)
MSPT, z‐score, median (IQR)
CST	−0.51 (−1.40, 0.08)
PST	−0.40 (−1.20, 0.35)
MDT	−0.57 (−1.89, 0.32)
WST	−0.69 (−2.05, 0.21)
PDDS, quartile, n (%)
first quartile (PDDS 0)	1126 (38.0)
second quartile (PDDS 1)	621 (20.9)
third quartile (PDDS 2–3)	691 (23.3)
fourth quartile (PDDS 4–7)	528 (17.8)

Abbreviations: CST, Contrast Sensitivity Test; IQR, interquartile range; MDT, Manual Dexterity Test; MS, multiple sclerosis; MSPT, Multiple Sclerosis Processing Test; PDDS, Patient‐Determined Disease Steps; PST, Processing Speed Test; SD, standard deviation; WST, Walking Speed Test.

3.2. Association of all Imaging Features With Disability Outcomes

Among the 93 radiomic features analyzed, 81 (87.1%) were statistically significantly associated with PDDS (first quartile [0−1] vs. fourth quartile [4−7]; Table 2). For MSPT, 37 (39.8%) were associated with WST, 33 (35.5%) with MDT, 11 (11.8%) with PST, and 5 (5.4%) with CST. For reference, of the eight key conventional MRI features analyzed, all were statistically significantly associated with PDDS, while six were associated with each component of MSPT.

TABLE 2.

Radiomic features associated with disability outcomes.

Subsets of radiomic features based on ability to differentiate between cohorts	Proportion of radiomic features showing significant (p < 0.05) association with disability outcomes
	PDDS	WST	MDT	PST	CST
Overall, n (%) n = 93	81 (87.1)	37 (39.8)	33 (35.5)	11 (11.8)	5 (5.4)
PwMS versus HCs, n (%) n = 12	10 (83.3)	6 (50.0)	6 (50.0)	1 (8.3)	1 (8.3)
Phenotype a , n (%) n = 35	35 (100)	31 (88.6)	26 (74.3)	11 (31.4)	5 (14.3)
Age b , n (%) n = 69	61 (88.4)	28 (40.6)	20 (29.0)	11 (15.9)	5 (7.2)
Race c , n (%) n = 69	58 (84.1)	24 (34.8)	19 (27.5)	5 (7.2)	2 (2.9)

Note: Each row corresponds to a subset of radiomic features shown to differentiate between participant cohorts. Each column represents the proportion of that subset of radiomic features that is significantly associated (adjusted p < 0.05) with each disability outcome, expressed as the number followed by the percentage of associated features in that subset. p‐Values calculated using linear regression for differences between pwMS and HCs (adjusted for age and sex), and for subgroups of pwMS (adjusted for age, sex, race, disease duration, MS course, and body mass index). The association of imaging features with disability outcomes was assessed using logistic regression (PDDS, in quartiles of first vs. fourth quartile) and linear regression (WST, MDT, PST, CST), adjusted for the above covariates.

Abbreviations: CST, Contrast Sensitivity Test; HCs, healthy controls; MDT, Manual Dexterity Test; PDDS, Patient‐Determined Disease Steps; PST, Processing Speed Test; PwMS, people with multiple sclerosis; WST, Walking Speed Test.

^a Relapsing versus progressive multiple sclerosis.

^b Age in quartiles (highest vs. lowest).

^c African American versus non‐African American.

3.3. Imaging Features in pwMS Versus HCs

Among the 93 radiomic features analyzed, 12 (12.9%) differed between pwMS and HCs (Figure 2A). Four of these features were of the GLCM class, all of which were also associated with disability as measured by PDDS, WST, and MDT. Each GLCM feature that was higher among pwMS was also higher among those with worse disability (i.e., higher PDDS and lower WST/MDT), and vice versa. Radiomic features belonging to other classes showed variable association with disability outcomes (Table 2). By comparison, the spinal cord area (SCA) was the only conventional MRI feature that was not different between pwMS and HCs. All seven conventional brain MRI measures analyzed were significantly different between pwMS and HCs, five of which were associated with PDDS and all components of MSPT, of which all features exhibited the same direction of difference between pwMS (vs. HCs) and those with worse disability (Figure 2B).

FIGURE 2.

Performance of all 93 radiomic features and key conventional brain (and upper cervical spinal cord) MRI features, respectively, in distinguishing pwMS versus HCs, and between pwMS by age (in quartiles; comparing first vs. fourth), disease course (comparing relapsing vs. progressive), and race (comparing non‐AA vs. AA), and in correlating with disability outcomes (PDDS [comparing first vs. fourth quartile] and MSPT components [by z‐score]). Green indicates a statistically significant difference/association (p < 0.05); red indicates a nonstatistically significant difference/association. “+” indicates a positive direction of association; “−” indicates a negative direction of association. For example, T2 lesion volume is higher in pwMS versus HCs, higher in progressive MS versus relapsing MS, higher in pwMS in the fourth quartile of age versus the first quartile of age, higher in AA versus non‐AA pwMS, higher in pwMS in the highest PDDS quartile versus the lowest PDDS quartile, and higher in pwMS with lower WST, MDT, PST, and CST z‐scores. p‐Values calculated using linear regression for differences between pwMS and HCs (adjusted for age and sex), and for subgroups of pwMS (adjusted for age, sex, race, disease duration, MS subtype, and body mass index). The association of imaging features with disability outcomes was assessed using logistic regression (PDDS) and linear regression (WST, MDT, PST, CST), adjusted for the above covariates. AA, African American; CST, Contrast Sensitivity Test; HCs, healthy controls; MDT, Manual Dexterity Test; MS, multiple sclerosis; MSPT, Multiple Sclerosis Processing Test; PDDS, Patient‐Determined Disease Steps; PMS, progressive multiple sclerosis; PST, Processing Speed Test; PwMS, people with multiple sclerosis; RR, relapsing remitting; WST, Walking Speed Test.

3.4. Imaging Features Across MS Disease Course

Thirty‐five of 93 (37.6%) radiomic features differed significantly between pwMS with a relapsing versus progressive course. The majority of these features were significantly associated with PDDS (100%), WST (88.6%), and MDT (74.3%), with 31.4% and 14.3% associated with PST and CST, respectively (Table 2). Among all 35 radiomic features, those that were higher among people with progressive MS were higher among those with worse disability, and vice versa. Similar to the pattern observed for features distinguishing pwMS from HCs, seven of these features were of the GLCM class, all of which were specifically associated with PDDS, WST, and MDT (Figure 2A). By comparison, four of eight conventional MRI features were different between pwMS with a relapsing versus progressive course; these four features were all associated with PDDS and MSPT. For all four features, higher values were associated with worse disability outcomes, and vice versa (Figure 2B).

Sixty‐nine (74.2%) radiomic features were significantly different between pwMS in the lowest (<38.3 years) versus the highest (>55.9 years) quartiles of age. The proportions of these features that were associated with disability outcomes were similar to those observed for all 93 radiomic features (Table 2). Where an association was seen with disability, the majority of radiomic features that were higher among the oldest age quartile were also higher among those with worse disability, and vice versa. Among these radiomic features differing by age, no clear pattern of association was observed between specific feature classes and disability outcomes (Figure 2A). All eight conventional MRI features differed between age quartiles of pwMS, with six associated with PDDS and MSPT (Figure 2B). Conventional MRI features that were higher among those in the oldest age quartile were also higher among those with worse disability, and vice versa.

Sixty‐nine (74.2%) radiomic features differed significantly among pwMS by race (AA vs. non‐AA), and these features exhibited similar associations with disability outcomes as compared to analyses involving all 93 radiomic features, with no clear pattern of feature involvement appreciable (Table 2 and Figure 2A). There was also no discernible pattern in the direction of difference of radiomic features by race (AA vs. non‐AA) and disability status (higher vs. lower disability). By contrast, only two of eight conventional MRI features (T1 and T2 lesion volume) differed by race, both of which were higher among AA pwMS, and were associated with higher PDDS and lower MSPT z‐scores (Figure 2B).

3.5. Imaging Features Most Strongly Associated With Disability

Of all disability outcomes analyzed, radiomic features consistently demonstrated the strongest correlation with PDDS score, followed by WST z‐score and MDT z‐score. Several classes of radiomic features were represented among those exhibiting the strongest association with disability outcomes (Figure 3A). The largest correlation coefficients with PDDS were observed for first‐order energy (ρ = −0.33; p < 0.001), GLRLM gray‐level nonuniformity (ρ = −0.29; p < 0.001), GLDM dependence nonuniformity (ρ = −0.28; p < 0.001), NGTDM coarseness (ρ = 0.26; p < 0.001), and GLSZM large area high gray‐level emphasis (ρ = −0.26; p < 0.001).

FIGURE 3.

Scatterplot matrix showing the correlation (Spearman's rho coefficients) between key radiomic features and conventional MRI features, respectively, and disability outcomes among pwMS. All p‐values < 0.05. CST, Contrast Sensitivity Test; GLDM, gray‐level dependence matrix; GLRLM, gray‐level run length matrix; GLSZM, gray‐level size zone matrix; MDT, Manual Dexterity Test; NGTDM, neighboring gray tone difference matrix; PDDS, Patient‐Determined Disease Steps; PST, Processing Speed Test; WST, Walking Speed Test.

By comparison, the strongest correlations with PDDS observed among conventional MRI features were for SCA (ρ = −0.26; p < 0.001) and whole brain fraction (ρ = −0.24; p < 0.001; Figure 3B). Of note, conventional MRI brain features exhibited the strongest correlations with PST—particularly T2 lesion volume (ρ = −0.35; p < 0.001), T1 lesion volume (ρ = −0.34; p < 0.001), and whole brain fraction (ρ = 0.34; p < 0.001).

4. Discussion

In this study, we identified radiomic features extracted from spinal cord MRI that distinguished pwMS from HCs. We observed features differentiating pwMS by relapsing versus progressive course, age, and race. The majority of radiomic features analyzed were associated with disability as measured by PDDS, with several features more strongly correlated with PDDS as compared to conventional MRI features. A relatively smaller number of radiomic features were associated with WST and MDT. A subset of radiomic features from the GLCM class distinguished pwMS from HCs and pwMS with relapsing versus progressive disease, and were significantly associated with PDDS, WST, and MDT. In our study, radiomic spinal cord features outperformed conventional spinal cord features (and conventional brain MRI features) in correlations with PDDS and WST, highlighting the value of radiomics derived from this difficult‐to‐image region.

New and evolving MRI techniques have enabled remarkably accurate visualization of the brain and granted insight into key pathologic processes in MS [3, 4]. However, brain MRI measures rarely explain overall disability in pwMS, particularly among those with high levels of disability [5, 6]. Disability progression in MS—particularly as it pertains to ambulation—is more closely aligned with spinal cord pathology, which is intuitive considering the spinal cord's small size and high concentration of structurally eloquent motor and sensory pathways [7, 8, 9, 44, 45]. Paradoxically, its small size—and artifacts due to respiration motion, cerebrospinal fluid flow, and cardiac pulsation in particular—have contributed to the comparative stagnation of spinal cord MRI technologies relative to brain MRI. Challenges in distinguishing subtle pathological changes from artifact on spinal cord MRI have led to an increasing reliance on brain MRI both in clinical practice and trials. As such, there has been greater incentive and focus on refining brain MRI rather than spinal cord MRI. This may partially explain why spinal cord MRI has not been observed to correlate with disability as strongly as may be expected among pwMS [7, 46, 47].

The application of radiomics to medical imaging analysis offers a potential solution to the challenges inherent in interpreting spinal cord MRI. Radiomics extracts predefined mathematical features from medical images. Information can be obtained based on signal intensities within the tissue microenvironment that may be below the limits of human visual detection, producing a mathematical phenotype characterizing ROI [11, 12]. This approach may overcome issues related to low resolution of spinal cord MRI and low sensitivity by visual inspection, as it reduces reliance on neurologists and radiologists to identify subtle pathological changes. In doing so, radiomics may uncover pathologic processes contributing to disability, in turn enhancing knowledge of disease mechanisms and potentially demonstrating better correlation with clinical outcomes.

To date, radiomics has primarily been utilized in oncology. However, it is being increasingly evaluated in neurological disease, and several studies have assessed its use in MS. The majority of these studies have explored its application in differentiating MS from other demyelinating conditions [13, 14, 15, 16, 17]. Radiomic features extracted from brain MRI have also been shown to distinguish pwMS from HCs [48]. Basic texture analysis of spinal cord MRI has demonstrated differences in gray‐level features between pwMS and HCs, but the performance of more advanced radiomic features in distinguishing pwMS from HCs has not yet been evaluated [49]. We identified radiomic features from several classes—predominantly GLCM and GLRLM—that differed between pwMS and HCs, therefore, grounding our radiomic findings in normative data. The relatively low proportion of radiomic features exhibiting significant differences between these cohorts may be related to the small size of the HC cohort (n = 41) in comparison to the MS cohort (n = 2966). Future studies with a larger cohort of HCs may be informative in this regard.

To our knowledge, spinal cord radiomic features have not been evaluated among specific subgroups of pwMS. In the present study, we demonstrate the ability of radiomics to differentiate between pwMS with relapsing versus progressive disease courses, suggesting that the features extracted reflect pathological processes operative at varying stages of the disease. The robustness of the association between these radiomic features and progressive clinical phenotypes is affirmed by the consistent association of these features with measures of global disability in a similar biological gradient. Interpretation of the direction of difference of radiomic features between subgroups is dependent on the feature in question. Higher‐order radiomic features measure textural differences within tissue, with greater heterogeneity thought to reflect more tissue damage in that region. Therefore, among radiomic features directly measuring pixel heterogeneity, higher values may reflect more severe myelin and axonal pathology, and vice versa for those measuring homogeneity [50]. The robustness of the observed clinical associations is further supported by the proportion of radiomic features distinguishing pwMS by age, considering that advanced age is associated with a progressive clinical course and more neurodegenerative pathology [51, 52]. Similarly, AA pwMS have been shown to exhibit higher rates of inflammation and accelerated neurodegeneration versus non‐AA pwMS [20, 53]. These pathological differences likely underpin the large number of radiomic features differing significantly between AA and non‐AA pwMS in the current study. Conventional brain MRI measures performed relatively poorly in this aspect, with only T1 and T2 lesion volume delineating these groups, but not conventional spinal cord or other brain measures. This suggests that spinal cord radiomic features may be more sensitive for neurodegenerative processes, distinguishing AA from non‐AA pwMS as compared to conventional brain and spinal cord MRI features.

Radiomic features extracted from brain MRI have shown an association with clinical disability among pwMS, both cross‐sectionally and in predicting disability progression as measured by Expanded Disability Status Scale (EDSS) [21, 22]. Basic texture features (mean gradient and mean gray level) in spinal cord MRI have previously demonstrated moderate correlation with EDSS (Spearman's rank correlation coefficient, r = 0.45 and r = 0.46, respectively) [49]. However, the association of more advanced spinal cord radiomic features with disability outcomes has not previously been assessed. In the present study, many radiomic features exhibited stronger correlations with measures of global disability as compared to conventional MRI features. While this is likely to be expected in comparisons of features extracted from the spinal cord versus the brain due to the topographical importance of the spinal cord to ambulation, several spinal cord radiomic features correlated more strongly with PDDS and WST as compared to the conventional measure of SCA. Upper cervical cord cross‐sectional area is among the most commonly utilized spinal cord imaging biomarkers in MS research, largely due to its robust correlation with global disability outcomes [54, 55]. Therefore, the stronger association of radiomic features with PDDS and WST suggests a potential role of spinal cord radiomic features in providing additional information for predicting disability in future clinical trials as compared to conventional spinal cord MRI.

A significant challenge in the application of radiomics is translating the mathematical output extracted into clinically meaningful data. Recognizing patterns of feature differences may facilitate greater insight into their nature and clinical significance. In the current study, we identified patterns of radiomic features differing between subgroups and predicting disability. For example, we observed a notable pattern of features of the GLCM class that were significantly different between pwMS and HCs, as well as between relapsing and progressive MS courses, and that were specifically associated with PDDS, WST, and MDT. However, GLCM features were not among those exhibiting the strongest overall correlation with disability; in fact, each of the five features with the strongest correlation coefficients for PDDS belonged to a different feature class. It should also be noted that the feature exhibiting the strongest correlation with PDDS, energy (ρ = −0.33), was generated from first‐order statistics. First‐order statistics describe the distribution of voxel intensities within an ROI, but do not describe spatial interactions and strongly depend on the number of bins selected within the region [56]. Therefore, first‐order statistics should not be interpreted in isolation without the context of higher‐order textural features. Further studies are needed to refine radiomics as it is applied to spinal cord MRI in MS. This may result in recognition of a more clearly defined and biologically based representation of radiomic features associated with clinical course and disability outcomes.

This study has several limitations. We elected to analyze radiomic features applied to a portion of the upper cervical cord included in brain MRI scans rather than dedicated spinal cord imaging. This was primarily to enable comparison of radiomic features with conventional brain MRI features, which, for reasons discussed previously, represent a “gold standard” of reliability and reproducibility relative to spinal cord MRI [3, 46, 47]. We quantified lesions within the brain only, rather than combined brain and spinal cord lesion volume, due to the low sensitivity of visual inspection of spinal cord MRI to detect lesions, as demonstrated in post‐mortem studies, as well as the differing associations of brain and spinal cord lesions with disability outcomes [57]. However, this decision limited direct comparison of clinical correlations between several conventional MRI and radiomic features. We elected to focus only on several preselected conventional MRI features as a representative sample, so as to maintain focus primarily on radiomic features. As such, detailed analysis of additional conventional MRI features differing between cohorts and correlating with disability outcomes could not be performed. Restricting the application of radiomics to the upper cervical cord may limit the generalizability of the findings to smaller, more mobile portions of the cord, where greater motion artifact may interfere with radiomic feature extraction. While upper cervical cord atrophy has been shown to reflect whole‐cord atrophy with quite high accuracy, it may not account for the regional variability of craniocaudal atrophy. In particular, lower cervical/upper thoracic cord atrophy may be more pronounced in progressive versus relapsing MS [58]. In such cases, restricting analyses to the upper cervical cord may underestimate the extent of whole‐cord involvement and weaken correlations with clinical outcomes. Therefore, although utilizing the spinal cord from brain MRI studies has efficiency implications when conducting routine clinical care imaging, further research is needed to examine the performance of radiomic features extracted from dedicated spinal cord MRI. Future studies, including one planned by our study group, will address this question by comparing radiomic and conventional features derived from dedicated spinal cord MRI, as well as evaluating the utility of radiomic features in predicting disability worsening longitudinally among pwMS.

Mean Dice scores for lesion segmentation were 0.72 and 0.69 in our validation and testing data sets, respectively. This indicates reasonably good performance for lesion segmentation based on prior studies, though slightly worse as compared to other MRI features analyzed (mean Dice scores ≥0.90) [59]. Despite this, T1 and T2 lesion volume still exhibited the strongest correlations with PST and MDT. Conversely, the relatively weak correlations of T1 and T2 lesion volume with PDDS and WST are unsurprising given that atrophy measures have been shown to predict global disability in MS more robustly than lesion volumetrics [2]. The high Dice scores of other MRI features analyzed suggest that coefficients between conventional MRI features capitulating neurodegeneration and PDDS/WST are accurate, and their performance in this aspect can be reliably compared with that of radiomic features. Finally, we included a relatively low number of HCs in our comparison cohort. However, our sample size of pwMS was far larger than those previously included in studies of radiomics in MS, strengthening the robustness of our findings.

In conclusion, we identified spinal cord radiomic features that distinguished pwMS from HCs, and subsets of pwMS by age, race, and disease course. Several radiomic features showed higher correlation with PDDS and WST than conventional MRI measures, including SCA. These findings suggest that the application of radiomics to spinal cord MRI may detect pathologic changes below the level of detection by traditional visual inspection, and that spinal cord radiomic features may have utility as biomarkers predicting disability in MS. Comparison of radiomic features with conventional MRI features is dependent on accurate techniques for segmentation and feature extraction. Further studies are needed to assess the performance of radiomic features extracted from lower, more mobile portions of the cord using dedicated spinal cord MRI scans. Additionally, studies across different data sets are needed to further ground radiomic features in clinical outcomes. This may facilitate a clearer understanding of the clinical significance of distinct radiomic features. The implementation of radiomics to spinal cord MRI may signal a transition from the search for data within the spinal cord to the search for meaning within the data extracted.

Disclosures

DO has received research support from the National Institutes of Health, National Multiple Sclerosis Society, Patient Centered Outcomes Research Institute, Race to Erase MS Foundation, Genentech, Genzyme, Bristol Myers Squibb, and Novartis, and consulting fees from Bristol Myers Squibb, Genentech/Roche, Novartis, and Contineum Therapeutics.

Conflicts of Interest

The authors declare no conflicts of interest.

Lambe J., Thompson N. R., Li Y., Nakamura K., and Ontaneda D., “Association of Spinal Cord Radiomic Features and Disability in Multiple Sclerosis.” Journal of Neuroimaging 35, no. 5 (2025): e70089. 10.1111/jon.70089