Exploratory SWI-based texture analysis of the substantia nigra on MRI in parkinson’s disease

Abstract

This study aimed to compare the substantia nigra of patients with Parkinson’s disease (PD) and healthy controls using magnetic resonance imaging texture analysis findings. Texture analysis was performed on the pixels within each region of interest in the substantia nigra for all participants. The entire imaging analysis workflow was implemented using an in-house algorithm developed using Matrix Laboratory. This study included 76 participants (37 patients and 39 controls). There were no significant differences between the groups in terms of age or sex (P > .05). In the unadjusted analysis, kurtosis, skewness, and uniformity differed significantly between patients and controls (P < .05); however, after correction for multiple comparisons, only kurtosis remained statistically significant. Receiver operating characteristic curve analysis demonstrated moderate discriminative performance, with kurtosis showing the most robust results. These findings suggest a potential supportive role of susceptibility-weighted imaging-based texture analysis in PD; however, given the exploratory design and the absence of external validation, the results should be interpreted with caution. The findings of this study suggest that texture analysis applied to imaging data of patients with PD may have a potential supportive role in the diagnostic evaluation of the disease.

1. Introduction

Cells that produce dopamine, a chemical that provides messaging between brain regions, are concentrated in certain cerebral regions called the substantia nigra. Symptoms of Parkinson’s disease (PD) are observed when 60% to 80% of these cells are lost, leading to dopamine deficiency.^[1,2] Intensive efforts are ongoing to develop therapeutic and diagnostic approaches aimed at slowing the progression of the disease.^[2–7]

Currently, PD can be diagnosed clinically. Neuroimaging is primarily used to exclude alternative conditions such as intracranial mass lesions, metabolic disorders, drug-induced parkinsonism, and normal pressure hydrocephalus in patients with suspected PD.^{[2–4,8–10]}

It has been reported that the substantia nigra is the most sensitive area that exhibits atrophy in the pars compacta when compared with other subcortical structures in the discrimination of patients with PD from healthy individuals.^[7,11–13]

In PD, early and reliable diagnosis largely relies on clinical evaluation, with imaging methods serving primarily a supportive role. Quantitative assessment of the structural and microstructural features of the substantia nigra may enable a more objective characterization of disease-related changes and support the development of imaging biomarkers. Despite increasing interest in magnetic resonance imaging (MRI)-based radiomics and texture analysis, data specifically addressing susceptibility-weighted imaging (SWI)-based texture characteristics of the substantia nigra remain limited.^[14,15]

Digital 2-dimensional images are commonly used in clinical practice and consist of small rectangular elements called pixels. Each pixel is defined by a 2-dimensional spatial coordinate and an associated gray-level intensity value. The image texture reflects the distribution and spatial arrangement of the gray-level values within a given region. Texture analysis comprises various mathematical and statistical methods used to quantify spatial variations in grayscale intensity, including signal characteristics, pixel relationships, gray-level density, and texture features that enable the objective assessment of tissue heterogeneity.^[16–19]

This study aims to evaluate SWI-based texture parameters of the substantia nigra using a region of interest (ROI)-based approach and to test the hypothesis that these parameters differ between patients with PD and healthy controls.

2. Materials and methods

The study was approved by the Firat University Non-Invasive Research Ethics Committee (February 21, 2019; Session: 04; Decision No: 15). All procedures performed in this study were conducted in accordance with the ethical principles of the Declaration of Helsinki. Informed consent was waived due to the retrospective nature of the study.

2.1. The study population

In this retrospective study, clinical and laboratory data were obtained from the Firat University Hospital Information System (HIS), and MRI images were retrieved from the picture archiving and communication system. The study included 37 consecutive patients diagnosed with and treated for PD at the neurology clinic between January 2017 and February 2019, as well as 39 age- and sex-matched control subjects randomly selected from individuals who underwent cranial MRI for other indications and had no cranial pathology. During the study period, 37 eligible PD patients were included in the archive. Exclusion criteria included a history of other neurological or psychiatric disorders, secondary parkinsonism, or atypical PD progression. None of the control subjects had a history of neurological diseases, hypertension, or diabetes.

2.2. Image acquisition/processing

MRI examinations were performed using a 1.5–Tesla Philips Ingenia system (Philips Healthcare, Best, The Netherlands). As this was a retrospective study, all MRI data were retrieved from the picture archiving and communication system archive and acquired using the institution’s routine, standardized cranial MRI protocol. Patients whose MRI examinations did not follow the routine cranial MRI protocol were excluded from this study. In the routine cranial MRI protocol, axial SWI was acquired using the following parameters: repetition time (TR), 51 ms; echo time (TE), 12 ms; echo spacing, 11 ms; flip angle, 20°; number of excitations (NEX), 1; 4 echoes; field of view (FOV), 230 × 185 mm; matrix, 272 × 184; slice thickness, 2 mm; 130 contiguous slices; and acquisition time, 3 minutes 26 seconds. The signal-to-noise ratio (SNR) settings were applied consistently across both patient and control examinations.

2.3. Image analysis

Axial SWI MRI images were exported in Digital Imaging and Communications in Medicine 3.0 format to a 27-inch iMac computer (Apple Inc., Cupertino). The ROI tool in OsiriX v4.9 imaging software (Pixmeo, Switzerland) was used to extract pixel values for texture analysis. A fixed ROI size was not predefined; instead, ROIs were manually drawn to include the entire substantia nigra, with the pars compacta and pars reticulata delineated separately on the axial images for each patient and control subject (Fig. 1). ROIs were defined by a radiologist with more than 10 years of neuroimaging experience, based on established anatomic atlases.^[20,21] The radiologist was not blinded to group assignment. Pixel values from each ROI were exported in XML (Extensible Markup Language) format to a Windows 10 computer (Microsoft Corporation, Redmond) and processed using Microsoft Excel 2013 (Microsoft Corporation, Redmond). Texture features, including mean gray-level intensity, standard deviation, minimum, maximum, median, variance, entropy, uniformity, skewness, kurtosis, and the percentage of pixels below 1 standard deviation (Size %L), above 1 standard deviation (Size %U), and between these thresholds (Size %M), were calculated using an in-house algorithm implemented in Matrix Laboratory, as described in the literature.^[17–19]

Figure 1.

A cross-sectional susceptibility-weighted imaging (SWI) brain image (A) selected for the placement of regions of interest (ROIs) in an individual, and the marked ROIs corresponding to the substantia nigra nucleus (B). ROI = region of interest, SWI = susceptibility-weighted imaging.

2.4. Statistical analysis

Statistical analyses were performed using the IBM Statistical Package for the Social Sciences Statistics for Windows (version 25.0; IBM Corp., Armonk). Data are presented as the mean ± standard deviation. Normality was assessed using the Kolmogorov–Smirnov test. Depending on the data distribution, between-group comparisons were conducted using the independent-samples t-test or the Mann–Whitney U test. A general linear model was used to adjust for age, sex, age at disease onset, and disease duration. Statistical significance was set at P < .05. Receiver operating characteristic (ROC) curve analysis was performed to evaluate the discriminatory performance of the texture parameters between the patient and control groups. Post hoc power analysis was conducted using PASS 11 software (NCSS, LLC, Kaysville). No data were missing.

3. Results

The mean age was 66.74 ± 14.57 in the patient group, while it was 66.24 ± 10.77 in the control group. There was no statistically significant difference between the groups in terms of age (P = .385).

Eighteen members of each group were female (48.6% and 46.2%, respectively), and there was no statistically significant difference between the groups based on sex (P = .828).

The mean age of disease onset was 58.1 ± 9.9 years, and ranged from 35 to 78 years, and the mean disease duration was 7.9 ± 3.4 years and ranged from 3 to 15 years.

Analysis of covariance (ANCOVA) showed that age, sex, age at disease onset, and disease duration did not affect the results (P > .05).

The comparison of texture analysis findings revealed that kurtosis and uniformity were significantly lower, whereas skewness was significantly higher, in the patient group compared with the control group in the unadjusted analysis (P < .05) (Table 1, Figure 2). After correction for multiple comparisons using the Benjamini–Hochberg false discovery rate (FDR) method, only kurtosis remained statistically significant (FDR-adjusted P = .026) with a moderate-to-large effect size (ES = 0.676). Skewness (ES = −0.614, FDR-adjusted P = .065) and uniformity (ES = 0.735, FDR-adjusted P = .065) demonstrated relatively large effect sizes but did not retain statistical significance after FDR correction. All other texture parameters showed nonsignificant adjusted p values (Table 1).

Table 1.

Distribution of values of analyzed parameters of the substantia nigra nucleus of the groups.

	Control (39)		Parkinson’s disease (37)		p
	Mean	S.D.	Mean	S.D.
Mean	317.35	108.74	327.81	133.02	.831*
Standard Deviation	38.94	14.91	48.19	58.37	.921*
Minimum	207.85	98.87	219.59	92.17	.594†
Maximum	408.00	134.71	445.99	250.28	.763*
Median	318.14	108.69	325.83	127.42	.876*
Variance	1732.66	1325.64	5636.55	23,677.21	.921*
Entropy	6.76	0.42	6.77	0.38	.979*
Size %L	15.85	2.01	16.80	2.41	.064†
Size %U	16.48	1.84	17.02	2.37	.270†
Size %M	67.67	3.25	66.18	3.71	.065†
Kurtosis	3.05	0.65	2.64	.56	.002 *
Skewness	−0.22	0.46	0.04	.36	.015 *
Uniformity	0.34	0.09	0.28	.06	.013 *

S.D.: Standard Deviation, %L: percentage of low-intensity pixels, %U: percentage of high-intensity pixels, %M: percentage of mid-intensity pixels.

^*Mann–Whitney U test.

^†Independent-Samples T-Test.

Figure 2.

Distribution of kurtosis values in patients with Parkinson’s disease and healthy controls.

ROC analysis demonstrated moderate discriminative performance for kurtosis. The area under the curve (AUC) for kurtosis was 0.702 (95% CI: 0.583–0.821; P < .001), with an optimal cutoff value of < 2.664, yielding a sensitivity of 71.8% and a specificity of 64.9% (Fig. 3). The statistical power to detect group differences for kurtosis at an alpha level of 0.05 was 83%.

Figure 3.

Receiver operating characteristic (ROC) curve illustrating the discriminative performance of kurtosis for distinguishing patients with Parkinson’s disease from healthy controls. The area under the curve (AUC) with 95% confidence interval and the optimal cutoff value are reported in the Results section. AUC = area under the curve, ROC = receiver operating characteristic.

Skewness showed an AUC of 0.663 (95% CI: 0.539–0.786; P = .010) with a cutoff value of > 4.818, while uniformity yielded an AUC of 0.667 (95% CI: 0.544–0.790; P = .008) with a cutoff value of < 0.322. Although skewness and uniformity demonstrated moderate discriminative performance, these parameters did not retain statistical significance after FDR correction and were therefore considered exploratory.

Discussion

In recent years, texture analysis methods have become prominent in neuroradiological analyses, mostly conducted on MRI images and the cerebral parenchyma. However, most studies have focused on malignancy diagnosis and typology, multiple sclerosis, and Alzheimer’s disease.^[19,22–25] Texture analysis studies of PD are relatively limited in the literature. Although texture analysis has been widely described as a quantitative approach for characterizing subtle signal heterogeneity in medical imaging, its application to PD, particularly for evaluating the substantia nigra, remains less well-established compared with other neuroradiological conditions.^[13,26]

PD is predominantly a disorder in older adults, with the incidence increasing markedly after the age of 65 years, and it is consistently reported to be more common in men than in women. Van Den Eeden et al demonstrated that PD incidence varies substantially by age and sex, with higher rates in older age groups and male predominance. This demographic pattern is also consistent with clinical evidence indicating that PD is typically diagnosed in late adulthood and that sex-related differences may influence the clinical presentation and disease course. In the present study, the mean age of the PD group was above 65 years and 51.4% of the patients were male, which aligns with the expected epidemiological distribution of PD and supports the representativeness of our study. Because age and sex may also affect neuroimaging-derived measurements and could act as potential confounders, we additionally controlled for these variables in the statistical analysis using a general linear model, thereby minimizing their influence on the observed texture-based differences between the groups.^[2,4,27]

PD is characterized by progressive neurodegeneration involving the substantia nigra, which plays a central role in the pathophysiology of parkinsonian motor symptoms. Therefore, detailed assessment of the substantia nigra has become a major focus of neuroimaging studies aimed at distinguishing PD from healthy individuals and other parkinsonian syndromes. Advanced MRI-based techniques, including SWI, diffusion-weighted imaging and quantitative susceptibility mapping (QSM), are increasingly being used to evaluate structural and microstructural changes within this region and to identify imaging biomarkers associated with PD. In this context, our study contributes by applying SWI-based texture analysis to evaluate the pars compacta and pars reticulata of the substantia nigra separately. This approach may facilitate the objective quantification of signal heterogeneity and provide additional imaging markers to support PD discrimination in routine clinical MRI.^{[2,5,8,10,13,26,28]}

In a previous study using QSM in patients with PD, mean gray-level intensity values within the substantia nigra were reported to be significantly higher in the PD group than in healthy controls. Although the standard deviation values also showed a slight increase in the patients, this difference was not statistically significant. These findings suggest that QSM may be sensitive to susceptibility-related signal alterations in the substantia nigra, whereas variability-based measures, such as standard deviation, may show more subtle changes and may be influenced by factors such as ROI delineation and imaging parameters.^[13] In the present study, mean gray-level values, standard deviation, and variance were higher in the PD group than in the control group; however, these differences were not statistically significant. These results are generally consistent with previous reports using SWI and QSM that have demonstrated that the susceptibility-related signal characteristics of the substantia nigra may differ between patients with PD and healthy individuals. QSM has been proposed as an alternative to SWI because it enables quantitative assessment of magnetic susceptibility, thereby providing a more objective measure of tissue susceptibility changes. Nevertheless, both SWI and QSM are influenced by similar underlying susceptibility effects, and subtle differences in the signal intensity distribution may be reflected in histogram- and texture-based parameters.^[13,29]

Entropy is a texture parameter that reflects image irregularity and is commonly used to quantify the heterogeneity in gray-level intensity distributions. In a previous QSM-based study focusing on the substantia nigra in PD, entropy values were reported to be significantly higher in patients than in healthy controls, suggesting increased tissue heterogeneity and altered susceptibility-related signaling characteristics. In the present study, entropy values tended to be slightly higher in the PD group than in the control group; however, the difference was not statistically significant. Differences between studies may be explained by differences in sample size, MRI acquisition protocols and hardware, and the applied post-processing methods, including ROI delineation strategies and the use of SWI versus QSM sequences. In addition, entropy measurements are known to be sensitive to variations in image quality and acquisition parameters that may contribute to the variability across studies.^[13,30–32]

Kurtosis, which describes the peakedness of the histogram distribution; skewness, which reflects the degree of asymmetry; and uniformity, which represents the homogeneity of gray-level intensities, are commonly used histogram- and texture-based parameters for characterizing gray-level distributions in medical images.^{[17–19,30,32]} These metrics have been widely applied to quantify subtle tissue heterogeneity and signal intensity variations that may not be readily apparent on qualitative visual assessment.^[30,32] Notably, these parameters were not evaluated in the previous QSM-based study by Li et al.^[13] In the present study, unadjusted analyses demonstrated differences in kurtosis, skewness, and uniformity between the PD and control groups; however, after correction for multiple comparisons, only kurtosis retained statistical significance, suggesting that this parameter may represent a more robust texture feature.^{[17–19,30,32]} ROC curve analysis demonstrated moderate discriminative performance of the evaluated texture parameters. These findings suggest that SWI-based texture features, particularly kurtosis, may have a potential supportive role in distinguishing patients with PD from healthy controls; however, given the modest AUC values, they should not be interpreted as standalone diagnostic markers.

Previous studies that applied texture analysis to compare brain regions known or presumed to be affected by disease with regions considered normal have demonstrated that texture features can reveal measurable differences in tissue characteristics. However, the direction and magnitude of these differences have been inconsistent across studies. This variability may be explained by differences in MRI hardware and field strength, acquisition parameters, image preprocessing steps, and the specific texture extraction methods used, as well as by variations in the ROI definition and segmentation strategies. Despite these methodological differences, the overall literature supports the potential value of texture analysis as a quantitative tool for detecting subtle structural and signal alterations beyond conventional visual assessment.^[18,33–35]

In conclusion, the findings of this study suggest that texture analysis of brain MRI in patients with PD may provide supportive, rather than definitive, information for the diagnostic evaluation of the disease. However, to overcome the limitations mentioned above and eliminate inconsistent findings, more comprehensive and prospective studies should be conducted with the simultaneous use of various imaging modalities used in the diagnosis of PD.

4. Limitations

This study has several limitations. The absence of clinical variables that may influence imaging findings (e.g., medication use), the relatively small sample size, and the retrospective design may have affected the results and introduced potential selection bias. In addition, interindividual variability in the extent and distribution of involvement across substantia nigra subregions may have influenced ROI-based measurements.

ROIs were delineated by a single experienced radiologist based on anatomical atlases; however, due to the retrospective design and reliance on single-observer segmentation, intra- and interobserver reliability analyses could not be performed, which constitutes a limitation in terms of reproducibility. Furthermore, the in-house MATLAB-based algorithm was not externally validated against a reference standard or benchmark dataset, which may limit the accuracy, robustness, and generalizability of the extracted texture features and warrants cautious interpretation of the results.

Despite these limitations, the use of a widely accessible sequence such as SWI represents a practical advantage, particularly given the limited number of MRI texture analysis studies in PD, and suggests potential clinical applicability of the proposed approach.

Author contributions

Data curation: Murat Baykara.

Methodology: Murat Baykara.

Project administration: Murat Baykara, Hanefi Yildirim.

Formal analysis: Hanefi Yildirim.

Validation: Hanefi Yildirim.

Writing – review & editing: Duran Tavut.

Writing – original draft: Murat Baykara.