Background

Parkinson’s disease (PD) is the second most prevalent neurological disease that often influences individuals over the age of 65 years with a variety of etiology and clinical manifestations [1, 2]. PD is specified by a group of motor symptoms that include Bradykinesia (characterized by movement involuntary disorder), muscle stiffness, rest shaking, defects in patient state, and gaiting [3]. In addition, PD may manifest several other non-motor symptoms such as apathy, sadness, psychosis disorder, constipation, and sleep disturbance [4,5,6,7]. The main pathological and etiological characteristics of PD are destruction of substantial nigra dopaminergic neuronal cell in the midbrain extrapyramidal pathway [3]. Investigations have revealed an atypical accumulation of cytoplasmic components, specifically Lewy bodies with α-synuclein, in the neostriatum of Parkinson's disease patients [8, 9]. In addition, iron accumulation in the brain, particularly in the neostriatum, is another significant hypothesis in the pathophysiology of PD [8, 10, 11]. Additionally, neostriatal microstructural changes have been reported in PD patients [8, 12, 13]. Moreover, previous studies have revealed that in PD the caudate nucleus, putamen, and temporal area of the hippocampal exhibit volume reduction over the time, determined by magnetic resonance imaging (MRI) [13].

In recent years, quantitative analysis of brain structures based on MRI has become one of the new method for evaluating neurological disorders; in particular, radiomics is a new quantitative method that extracts a large number of features from medical images using data characterization algorithms [14]. These features have the potential to reveal specific pathological-based patterns and characteristics that fail to be appreciated by the naked eye [14]. Radiomics has recently garnered significant interest in Parkinson’s disease, as it has in other neurological diseases. A two-year longitudinal MRI-based study reported that there is a significance difference in structure of brain between PD and HC group. Moreover, the pattern of brain structure variations in PD patients correlated with clinical evaluation scores indicating disease extent [14, 15]. Rahmim et al. found that textural radiomic features of the caudate, putamen, and ventral striatum in dopamine transporter single-photon emission computed tomography (DAT SPECT) images can aid in predicting the PD progression model [16]. Several another previous studies focused on DAT SPECT radiomics, while another study fused SPECT and MRI images, implementing segmentation on MRI images and feature extraction on SPECT images [16, 17]. Another study suggests that standard structural MRI can predict Parkinson's disease progression. It highlights a simple radiomics signature based on whole brain white matter for tracking PD development, with sensitivity at 0.960, specificity at 0.600, and an accuracy of 0.854 [18].

Although radiomics studies have been conducted to exploring of MRI biomarkers associated with Parkinson's disease progression, there is still no comprehensive consensus on the results of radiomics and how to use it. In order to introduce robust radiomics biomarker that can aid in the diagnosis of the progression of Parkinson's, it is crucial to undertake further longitudinal research in the field of radiomics. This is particularly important because identifying the progression of Parkinson's disease caused to pose significant challenges for researchers and clinicians alike. More in-depth investigations into these MRI biomarkers are necessary to enhance diagnostic accuracy and improve patient outcomes. In this regard, this longitudinal study aimed to identify Parkinson's disease (PD)-specific radiomics biomarkers by analyzing textural radiomics features in caudate and putamen over two-year follow-up MRI images of PD subjects, compared to healthy control group.

Methods

Participants

Longitudinal 3 T MRI images were selected from the Parkinson's Progression Markers Initiative (PPMI) database (https://ida.loni.usc.edu/home/projectPage.jsp?project=PPMI). A total of 20 PD patients and 18 healthy control (HC) in age range of (50 −70) were selected. We included only samples that have normal cognitive condition or have not mild cognitive impairment (NO MCI) as well as at least two longitudinal MRI image series during two first years (baseline and follow-up). For all PD samples, follow-up MRI images were two years after baseline MRI; however, for HC group, eight subjects have follow-up MRI one year after baseline and ten subjects had MRI two years after baseline. Also we selected subjects that have same MRI pulse sequence across baseline and follow-up images. With these inclusion and exclusion criteria, only 18 healthy subjects with two-time-point follow-up MRI images were found from the database.

MRI imaging protocol

Three-dimensional magnetization prepared rapid gradient echo generalized (3D-MPRAGE) sequence of images was selected. 3D-MPRAGE is one of the most prominently used sequences for morphological brain imaging in clinical and investigating contexts; with a brief scan duration, the sequence offers good spatial resolution and acquires strong tissue contrast, covering the entire brain [19]. All of the MRI images have the same scan parameters as the following: TR = 2300 ms, TE = 3 ms, FA = 9°, slice thickness = 1 mm, Matrix X = 240 mm, Matrix Y = 256 mm, Matrix Z = 176 mm, TI = 900 ms, pixel spacing = X = 1.0 mm, Y = 1.0 mm, Z = 1.0 mm, 176 slices pulse sequence = gradient echo/inversion recovery, weighting = T1.. We applied 3D-MPRAGE sequence with generalized autocalibrating partially parallel acquisition (GRAPPA) technique in 3-Tesla MRI Siemens for all of the PD and HC individuals.

Region of interest delineation and feature extraction

Radiomics feature extraction was carried out by LIFEX software version (V.7.1.0) https://lifexsoft.org/index.php/downloads). LIFEX is an open-source and widely used software that first enables the user to complete the manual object delineation before switching to feature extraction [20]. Based on previous study, to reduce image noise image grayscale intensity level in LIFEX software was set to 32 bins [18]. In this study, 3D-ROI of bilateral caudate (CU) and putamen (PU) structures was delineated. Though these structures are bilateral in the brain [21, 22], these structures were delineated on the right and left brain side. A neuroradiologist with 10 years of experience (AJ) manually delineated regions of interest (ROIs) on caudate and putamen in axial MRI images. To mitigate boundary effects on radiomics results, the ROI was delineated to avoid close proximity to adjacent structures (Fig. 1). To minimize the influence of ROI size on radiomics feature amounts, consistent size of ROIs was delineated in all images [8, 23].

Fig. 1
Fig. 1
Full size image

Delineation of ROI was inside the structures and didn’t overpass the boundaries

Statistical analysis

Statistical analysis of covariance (ANCOVA) was carried out using SPSS V27.01 (IBM, SPSS, Inc., Chicago, IL, USA). At first, an analysis of covariance was used between HC subgroups (one-year follow-up and two-year follow-up), to determine two groups of HC subjects (1-year follow-up with 8 samples and two-year follow-up with 10 samples) that show significant change; if radiomics features did not significantly change between these two subgroups (1-year follow-up with 8 samples and two-year follow-up with 10 samples) over time, we could combine them into a single HC group with a sample size of 18. Afterward, analysis of covariance was also used between radiomics features of PD group and HC group. During ANCOVA, baseline radiomics features were used as covariates and follow-up features as dependent variables. This study employed a statistical significance threshold of P < 0.05.

Results

Demographic information of the subjects is summarized in Table 1. At first, the analysis of covariance between HC subgroups (one-year follow-up and two-year follow-up) showed that one-year HC group with 8 samples and two-year HC group with 10 samples didn’t show significant change (P > 0.05), so we combined the two groups in one group, and then, we have 18 individuals for healthy controls. A total of 32 s-order features were extracted from bilateral caudate and putamen. Among these 32 features between PD group and HC group, 18 features showed meaningful alteration; and their changes in the trend over time are listed in Table 2. These significant features were included of 7 Gy-level co-occurrence matrices (GLCM), 5 Gy-level run-length matrices (GLRLM), 2 neighboring gray-level dependence matrices (NGLDM), and 4 Gy-level zone length matrices (GLZLM). These 18 significant features showed meaningful change between the Parkinson's and healthy groups over two-year follow-up. Left caudate exhibited 12 significant features, surpassing the left/right putamen and right caudate, and its significance levels were also considerably higher [p = 0.001 − 0.008]. Almost all of the features indicate a kind of homogeneity and non-uniformity of gray-level pattern of voxels in images. A brief definition of textural radiomics features that be significant in our study is listed in Table 3.

Table 1 Demographic information of the study subjects
Full size table
Table 2 Mean value of significant features in HC and PD groups at baseline and follow-up; ANCOVA p-value; and trend of their changes during PD progression
Full size table
Table 3 Definition of textural radiomics features according to LIFEX platform
Full size table

Discussion

Our longitudinal study aimed to assess MRI-based radiomics feature changes at putamen and caudate in Parkinson's disease patients over the time. According to the significant features (Table 2), radiomics features in the left caudate showed a greater number and a stronger significance level than the results in the right structures. Our result slightly is in consistent with previous studies that the left side of the brain is often affected severely than the right side in Parkinson’s disease. Claassen D,O et al. reported nigrostriatal pathway revealed considerable degeneration on the left side of brain in Parkinson’s disease [24]. Yang Wenyi et al. showed that brain alteration such as atrophy or reduction in specific structures involved more the left hemisphere of Parkinson’s disease brain [25].

Iron and α-synoclein accumulation in nigrostriatal pathway may cause the alteration of ratio of iron and α-synoclein zones in the ROI in images [10,11,12,13] resulting in textural radiomics feature changes that differentiate PD from HC. In Parkinson's disease, iron accumulation in the substantia nigra causes microstructural changes undetectable by standard MRI analysis. Guangwei Du et al. used quantitative susceptibility mapping to evaluate nigral iron accumulation in Parkinson's disease (PD) from diagnosis to late stages. The study compared iron levels in substantia nigra, caudate, putamen, red nucleus, and globus pallidus of PD patients with disease durations of < 2 years, 2–6 years, and > 6 years. The findings suggest that iron accumulation is lower before dopaminergic medication, increases during disease progression, and eventually plateaus in the end stage [26]. Previous study proved that microstructural change occurred in different structures of Parkinson’s brain structures and QSM images also reveal this process and this is in consistent with our result that some features showed meaningful change in basal ganglia structures. Taisuke Harada et al. reported magnetic susceptibility in the basal ganglia (caudate, putamen, globus pallidus), substantial nigra, and red nuclei increased that this signs are consistent with parkinsonism so quantitative susceptibility map (QSM) images could be an accurate instrument for analysis of iron deposition in Parkinson’s disease [27].

Radiomics has been recently gained prominence in neurodegenerative studies, especially for mining of the pattern of microstructural degenerations [14]. Several previous studies have indicated that radiomics textural features reflecting heterogeneity can effectively differentiate PD patients from healthy controls [8]. Previous studies indicate an ascending pattern of iron accumulation in different brain structures during Parkinson's progression, resulting in an increase in the susceptibility-weighted imaging (SWI) in different structures [28,29,30]. Our results similar to previous studies (QSM and SWI technique) indicate a kind of homogeneity or uniformity pattern of gray level or gray-level runs for voxels; GLCM, GLRLM, NGLDM and GLZLM (Table 3). These radiomics feature alteration during Parkinson’s progression may be due to in the iron accumulation in that structures or we can say that iron accumulation caused alteration in homogeneity and non-uniformity and gray-level differences in these structures. The gray-level co-occurrence matrix (GLCM) calculates textural indices based on the spatial relationships between pairs of voxels. The neighborhood gray-level difference matrix (NGLDM or NGTDM) quantifies the gray-level differences between one voxel and its 26 neighbors in 3 dimensions, the gray-level run-length matrix (GLRLM) gives the size of homogeneous runs for each gray level, and the gray-level zone length matrix (GLZLM or GLSZM) quantifies the size of homogeneous zones for each gray-level in 2 dimensions or 3 dimensions [31]. In Table 4, results of recent several previous longitudinal and cross-sectional radiomics-based studies in Parkinson have been summarized and are compared with our findings. Although features with different names seem to be reported in previous studies, their underlying concepts are related to the heterogeneity or non-uniformity of gray levels.

Table 4 Results of previous studies that showed meaningful change during Parkinson’s disease and differentiation between Parkinson’s disease and healthy control
Full size table

Our study has several limitations. We extracted and analyzed only the second-order textural features, but other high-order features such as wavelet features may have advantages. We also used only LIFEX, a widely used software to extracted radiomics features; future studies can evaluate inter-software reproducibility by repeat such study by other software. Small sample size due to lack of longitudinal MRI images in PPMI database for healthy control subjects is another study limitation. Further research may be useful if such study develops deep learning to predict PD progression by extracting radiomics features. We delineate ROI manually; future studies with fully automatic segmentation may improve reproducibility and robustness of radiomics results [34].

Conclusion

Our radiomics analysis revealed that several textural radiomics features in caudate and putamen changes significantly during Parkinson’s disease progression. The left caudate exhibited a greater number of these features and a higher level of significance. Our radiomics features that based on a quantification of the pattern of gray level or gray-level runs of voxels, reflecting the heterogeneity and uniformity of pixel gray scales can provide diagnostic information undetectable by naked eye. Significant changes of radiomics textural features in left caudate may can be reflecting the underlying pathophysiology and therefore can serve as a noninvasive biomarker in disease management and treatment.