Texture analysis (TA) is an emerging subfield of radiomics, representing a non-invasive and objective method for the quantitative assessment of medical images. Texture features (TFs) can be used to quantitatively characterize image properties, such as uniformity, heterogeneity and randomness, as well as repetitive image patterns. TA has the potential to enable the extraction of additional diagnostic, predictive and prognostic information beyond what is visually perceptive (1). Traditionally, applications of TA include neuroimaging, musculoskeletal and oncological imaging, e.g. to assess tumor heterogeneity (2, 3). TA has been performed across different anatomical regions and modalities, with CT being the modality used most frequently, arguably due to its multitude of clinical applications, broad availability and data quality (4). Overall, musculoskeletal applications have been based on radiographic, magnetic resonance imaging (MRI) (5, 6) and CT data (7–9).

Osteoporosis is a systemic disorder of bone metabolism characterized by reduced bone mineralization and microarchitectural deterioration of osseous tissue (10, 11). Bone health and the closely associated vertebral fracture (VF) risk are dependent on bone mineral density (BMD) and bone microstructure which is primarily defined by the three-dimensional (3D) trabecular bone architecture as suggested by high correlations with micro-CT as reference (12, 13). Volumetric BMD (vBMD) can be measured on a vertebra-specific level with high spatial resolution using quantitative CT (QCT), while CT-based TA enables the evaluation of trabecular bone microstructure. Therefore, osteoporosis represents one of the most promising clinical applications of TA. In a small study cohort, Mookiah et al. showed that TA is feasible for opportunistic osteoporosis screening and can discriminate subjects with and without VFs accurately while exhibiting acceptable long-term scan-rescan reproducibility. However, the size of the used study population does not enable the derivation of TF reference values (7). Muehlematter et al. used machine learning algorithms on CT-based TA to better predict incident VFs (8). Through the combination of 3D-TA and regional vBMD, Valentinitsch et al. could improve the classification of prevalent VFs. In this study, the eventually analyzed TFs were selected from a variety of features, using an exponential search based on the Gini Index and subsequently used for classification of VFs which represent the most relevant clinical outcome of osteoporosis. However, the actual TF values were not reported (9). In spine imaging, commonly used TFs include first-order, second-order and higher-order statistical features, and reliable extraction of specific TFs was derived from sagittal reformations of up to 3 mm slice thickness in routine abdominal multi-detector CT (MDCT) scans with administration of intravenous contrast medium (IVCM) (7). However, TA for the evaluation of the osseous microstructure of the spine and ensuing bone health has not been standardized, yet.

One essential and time-consuming part of the TF extraction process is the segmentation to create individual regions of interest (ROIs) of the vertebral bodies, which has usually been performed manually on single slices. However, recent advancements in deep learning, in particular in convolutional neural networks, enable a standardized 3D segmentation by a fully automated pipeline (14–16). In order to discriminate TF values indicative of osteoporosis from normal values, and to enable longitudinal analyses of vertebral microstructure, the characterization of reproducible TF values is needed and can therefore be considered a prerequisite for diagnostic and therapeutic decisions in osteoporosis that use TFs.

Therefore, purposes of our study were (i), to identify CT-based TFs of vertebral bone with high long-term reproducibility which would be particularly important for longitudinal studies in the context of intervention monitoring, and (ii), to generate an automated standardized pipeline for segmentation and TA of the spine to establish characteristic TF values and determine physiological variations of the identified highly reproducible TFs with regard to gender, age and vertebral level in a mainly older adult population.

The present study performed automated spine segmentation and TF extraction of vertebral bodies in routine abdominal MDCT scans in a large-scale patient cohort. In a preceding analysis, long-term reproducibility of a set of commonly utilized TFs was evaluated. Six TFs with high reproducibility were identified and selected for further analysis. To the best of our knowledge, this is the first study combining fully automated segmentation, vBMD mapping and long-term reproducibility analysis to establish gender-, age-, and region-specific characteristic values and variations of vertebral TFs. This data could serve as foundation for a reliable differentiation of values indicative of osteoporosis from normal values across genders, age groups and spine regions. In particular, it could be a prerequisite for the longitudinal comparison of TFs.

Variance_global and Entropy showed a significant increase in cranio-caudal direction along the spine for both males and females. The other four TFs, which are all GLRLM-derived features, also showed significant region-dependent differences, except for females between the lower thoracic and lumbar spine. Five of the six TFs showed a significant difference between genders. None of the TFs showed a significant age-dependence. vBMD showed a significant decrease with age as well as statistically significant differences between genders and spine regions, confirming results from previous studies (13, 18, 29).

Dimension reduction and feature selection is a commonly performed step in TA studies to reduce redundancy, and to ensure that selected features fulfill certain conditions and have high relevance. However, the applied approaches are manifold and vary substantially between studies. Among many other methods, GLCM- and GLRLM-derived features can be restricted to or averaged across directions and distances (7, 30). Another popular approach is to only select reproducible features with high inter- and intra-reader agreement (8, 30). Mannil et al. additionally performed correlation analysis and excluded highly correlating TFs to reduce redundancy (30). The emergence of machine learning has further extended the repertoire of feature reduction methods. Using a random forest classifier to identify vertebral fractures, Valentinitsch et al. optimized the number of features based on the Gini importance and classification performance in an exponential search (9). We chose a more clinically driven approach and selected TFs based on a preceding long-term reproducibility analysis to identify features particularly suited for longitudinal comparisons.

Our study determined characteristic TF values based on CT images of vertebral bodies. The found values for Variance_global show substantial differences to the values reported by Mannil et al. (30), who performed the only similar study we could find in previous literature. This discrepancy, however, is potentially due to the limited comparability of the two studies. While Mannil et al. used manually prescribed ROIs of mid-sagittal planes of the vertebral body for two-dimensional (2D) TA (30), we utilized 3D-ROIs of the entire volume of the trabecular compartment of the vertebral body with subsequent 3D-TA. As a result, number and intra-vertebral localization of the included voxels as well as their gray value distribution vary remarkably between the two studies. Differences in CT image acquisition and reconstruction as well as slice thickness (2 mm vs. 3 mm) may additionally contribute to the discrepancy of the reported TF values. Fully automated labeling and 3D-segmentation of the spine represent major improvements in terms of time efficiency, reproducibility and utilization of available CT data when compared to manually performed 2D-segmentations. Therefore, the segmentation approach used in the present study should be considered as the standard for TA of the spine.

Characteristic values of TFs should be used and compared with great care and ideally only if acquisition and reconstruction parameters as well as postprocessing methods are identical. Therefore, we used images acquired at the same MDCT scanner with a standardized protocol, a fully automated and standardized segmentation method, as well as careful preparation of CT data prior to TF computation, including isotropic resampling and gray-level normalization.

The TFs Variance_global and Entropy showed significant regional differences, increasing from the mid-thoracic to the lumbar spine, irrespective of gender. Along the spine, more inferior vertebral bodies have larger size as well as a higher prevalence and degree of degenerative changes (29, 31). This could be potential explanations for the cranio-caudal increase in Variance_global since it reflects the spread of gray-level distribution. The GLCM feature Entropy is a measure of randomness in an image. However, it is not trivial to define what image properties it actually represents. According to Haralick et al., Entropy is associated with gray-level number, complexity and degree of structure in an image (21, 32). Therefore, the found region-dependent variations could indicate a gradient in architectural complexity along the spine, potentially suggesting variability in trabecular microstructure and skeletal integrity (33).

The TFs SRE, LRE, RP and RLN also showed gender-specific differences, as well as region-dependent differences, irrespective of gender for most features. These GLRLM features are highly associated with each other and, according to their defining equations, reflect the predominating length of gray-level runs, the distribution of gray-level run lengths, as well as the extent of linear structures (22). The variations we found may therefore indicate gender- and region-specific variability in trabecular microstructure, in line with previous ex vivo results (34).

There are several limitations to our study. First, we performed a retrospective analysis with limited availability of weight and height as well as other bone health-related data of the study cohort. Therefore, we did not adjust our results for variations in body mass index (BMI) or other potential confounders, such as smoking status, which are known to be correlated with BMD and could also have an influence on TFs. Second, n = 376 is relatively small sample size for establishing reference values, and the age range of the study is limited, in particular the age group under 40 years is not represented in our cohort. However, primary osteoporosis, as the most important condition for clinical application of the herein used approach and derived characteristic TF values, usually does not start before the age of 40 years. In contrast to values derived from a population of young individuals with normal bone health, age-, gender- and region-specific may also represent a viable approach for the differentiation of physiological and pathological bone alterations of the spine. Nevertheless, extending the analysis to an increased study population with a broader age range should be an aim of future investigations, in order to firmly establish a reference database. Third, the initial set of TFs we considered in our analysis was limited, and additional TFs (e.g. based on local binary patterns, wavelets or other transforms) could have been used (35, 36). Fourth, to maximize comparability of the computed values, CT data acquisition and postprocessing as well as TA were performed with homogenous and standardized settings. The restriction to a specific CT scanner and acquisition protocol may reduce the utility of the findings to other investigations using different settings. However, TF values, in particular for GLCM- and GLRLM-derived features, can only be reliably used for comparisons in studies with identical settings.

In conclusion, we established highly reproducible TF values for CT-based 3D-TA of vertebral bodies in a predominantly older study population, using an automated segmentation and quantification pipeline. These characteristic TF values and the found gender-, age-, and vertebral-level-specific variations can be considered a foundation for the reliable differentiation of physiological and pathological alterations and may be particularly important for future longitudinal studies.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by Ethikkommission der Technischen Universität München. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

MD wrote the manuscript. MD organized the database. MD performed the formal and statistical analysis. MD contributed to the concept and design of the study. MD developed major parts of the used software routines. NS, MH, AS, ML, CZ, JK, KS, and TB critically revised the manuscript. MH, AS, and ML contributed to the development of the software used in the study. TB conceptualized and designed the study. CZ, JK, KS, and TB provided scientific supervision and resources. JK and TB provided funding. TB contributed to the generation and organization of the database. All authors contributed to the article and approved the submitted version.

The present work was supported by the German Research Foundation (Deutsche Forschungsgemeinschaft, DFG, project 432290010, JK and TB), the German Society of Musculoskeletal Radiology (Deutsche Gesellschaft für Muskuloskelettale Radiologie, DGMSR, MD and NS), and the B. Braun Foundation (project BBST-D-19-00106, NS).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.