MRI texture analysis of femoral neck: Detection of exercise load-associated differences in trabecular bone

Authors

  • Lara C.V. Harrison MSc, MD,

    Corresponding author
    1. Tampere University Medical School, Tampere, Finland
    2. Department of Biomedical Engineering, Tampere University of Technology, Tampere, Finland
    3. Medical Imaging Centre, Tampere University Hospital, Tampere, Finland
    • Tampere University Hospital, Teiskontie 35, PO Box 2000, FIN-33521 Tampere, Finland
    Search for more papers by this author
  • Riku Nikander PhD,

    1. Bone Research Group, UKK Institute for Health Promotion Research, Tampere, Finland
    2. Science Center, Pirkanmaa Hospital District, Tampere, Finland
    3. Department of Medicine, University of Melbourne, Melbourne, Australia
    Search for more papers by this author
  • Minna Sikiö MSc,

    1. Department of Biomedical Engineering, Tampere University of Technology, Tampere, Finland
    2. Medical Imaging Centre, Tampere University Hospital, Tampere, Finland
    Search for more papers by this author
  • Tiina Luukkaala MSc,

    1. Science Center, Pirkanmaa Hospital District, Tampere, Finland
    2. Tampere School of Public Health, University of Tampere, Tampere, Finland
    Search for more papers by this author
  • Mika T. Helminen MSc,

    1. Science Center, Pirkanmaa Hospital District, Tampere, Finland
    2. Tampere School of Public Health, University of Tampere, Tampere, Finland
    Search for more papers by this author
  • Pertti Ryymin PhD,

    1. Medical Imaging Centre, Tampere University Hospital, Tampere, Finland
    Search for more papers by this author
  • Seppo Soimakallio MD, PhD,

    1. Tampere University Medical School, Tampere, Finland
    2. Medical Imaging Centre, Tampere University Hospital, Tampere, Finland
    Search for more papers by this author
  • Hannu J. Eskola PhD,

    1. Department of Biomedical Engineering, Tampere University of Technology, Tampere, Finland
    2. Medical Imaging Centre, Tampere University Hospital, Tampere, Finland
    Search for more papers by this author
  • Prasun Dastidar MD, PhD,

    1. Tampere University Medical School, Tampere, Finland
    2. Department of Biomedical Engineering, Tampere University of Technology, Tampere, Finland
    3. Medical Imaging Centre, Tampere University Hospital, Tampere, Finland
    Search for more papers by this author
  • Harri Sievänen ScD

    1. Bone Research Group, UKK Institute for Health Promotion Research, Tampere, Finland
    2. Science Center, Pirkanmaa Hospital District, Tampere, Finland
    Search for more papers by this author

Abstract

Purpose:

To assess the ability of co-occurrence matrix-based texture parameters to detect exercise load-associated differences in MRI texture at the femoral neck cross-section.

Materials and Methods:

A total of 91 top-level female athletes representing five differently loading sports and 20 referents participated in this cross-sectional study. Axial T1-weighted FLASH and T2*-weighted MEDIC sequence images of the proximal femur were obtained with a 1.5T MRI. The femoral neck trabecular bone at the level of the insertion of articular capsule was divided manually into regions of interest representing four anatomical sectors (anterior, posterior, superior, and inferior). Selected co-occurrence matrix-based texture parameters were used to evaluate differences in apparent trabecular structure between the exercise loading groups and anatomical sectors of the femoral neck.

Results:

Significant differences in the trabecular bone texture, particularly at the superior femoral neck, were observed between athletes representing odd-impact (soccer and squash) and high-magnitude exercise loading (power-lifting) groups and the nonathletic reference group.

Conclusion:

MRI texture analysis provides a quantitative method for detecting and classifying apparent structural differences in trabecular bone that are associated with specific exercise loading. J. Magn. Reson. Imaging 2011;. © 2011 Wiley Periodicals, Inc.

BESIDES CORTICAL GEOMETRY and thickness, trabecular architecture is also an essential determinant of whole-bone strength (1, 2). Specific microlevel organization of trabecular bone, i.e., length, orientation, and connectivity of bony rods and plates, accounts for the mechanical stiffness and strength of a given bone 3). Ultimately, the structure provides the whole bone with the capability to deform and absorb energy, which helps to cope with locomotive loading (4). Different imaging modalities with various analysis techniques have been successfully applied in both in vitro and clinical in vivo settings to characterize bone geometry and internal structure (5–10). Trabecular bone architecture is generally described by morphological parameters, such as bone volume/total volume (BV/TV), trabecular thickness (Tb.Th), trabecular separation (Tb.Sp), and trabecular number (Tb.N) (11, 12).

Bone texture is an important pattern property of images that has been quantitatively analyzed by radiography, computed tomography (CT) and MRI (13–16). Texture analysis (TA) refers to several mathematical methods to describe gray level dependence between pixels within the image. The co-occurrence matrix-based method (17, 18) has been shown to be useful for characterizing various study materials. Texture parameters are used to characterize the underlying structure of target tissues present in the given image.

In the clinical context, imaging techniques that do not involve invasive bone biopsies or radiation exposure would likely become first line methods. In light of these demands, out of several imaging modalities applicable for trabecular architecture characterization, MRI is free from ionizing radiation and thus a worthwhile option for trabecular assessment (19). However, present clinical MRI systems equipped with standard coils have limited capability to produce images with sufficient spatial resolution for the accurate separation of single trabeculae. Also, the length of the imaging session must be kept relatively short for patient comfort and avoidance of movement artifacts. Therefore, only low-resolution bone images can be obtained in the general clinical setting (5). The texture parameters can describe inherent gray level dependences in the image, and in case of bone image, these parameters reflect the structure of bone tissue within the limits of imaging resolution. In a recent study, wrist MR images were obtained with a 3 Tesla (T) scanner with isotropic resolution of 0.200 mm. These images and sub-sampled lower resolution images at 0.350, 0.500, and 0.700 mm were then analyzed. In these preliminary results, the co-occurrence matrix-based texture values of trabecular bone from lower resolution MR images correlated significantly with structural parameters (BV/TV, Tb.Sp. and Tb.N) extracted from the high-resolution MR images (20). MRI texture analysis based on histogram, run-length matrix, and co-occurrence matrix parameters have also been shown to characterize osteoporotic appearance in trabecular bone (13).

In this study, we applied the MRI texture analysis based on co-occurrence matrix-derived parameters, which have been successfully applied for quantitative image analysis both in diagnosis and prediction of outcomes in several medical conditions (21–23). Hip fractures pose a serious public health problem (24), and bone structure undoubtedly accounts for hip fragility (1, 2). Exercise, in turn, is known to be associated with reduced risk of hip fracture (25). Thus, the primary purpose of the present study was to evaluate whether the MRI texture analysis could detect potential differences in the femoral neck trabecular texture which may be attributable to long-term exercise loading. Previously an association of trabecular bone with high impact activity has been detected using morphological parameters obtained with 7T MRI (26) and from 1.5T MRI (27) of knee region. However, the present study comprises a greater number of athletes representing different exercise-loading types and their healthy referents in addition to imaging of clinically more relevant femoral neck. Our second purpose was to explore potential differences in trabecular texture at different anatomical sectors of the femoral neck cross-section.

MATERIALS AND METHODS

Ninety-one adult female athletes competing actively at the national or international level and 20 nonathletic referents participated in this cross-sectional study. The athletes were triple-jumpers (N = 9), high-jumpers (N = 10), soccer-players (N = 10), squash-players (N = 10), power-lifters (N = 17), endurance runners (N = 18), and swimmers (N = 18). They were recruited through national sports associations and local athletic clubs. The nonathletic, clinically healthy 20 referents were mainly students of the local nursing school. The Tampere University Hospital District Ethics Committee approved the study, and participants gave informed consent in writing.

The exercise-loading types represented by the athletes were grouped into five categories according to our protocol (28, 29): (i) the high-impact (H-I) exercise-loading group comprised triple-jumpers and high-jumpers; (ii) the odd-impact (O-I) exercise-loading group comprised soccer and squash players; (iii) the high-magnitude (H-M) exercise-loading group comprised power-lifters; (iv) the repetitive, low-impact (L-I) exercise-loading group comprised endurance runners; and (v) the repetitive, nonimpact (N-I) exercise-loading group comprised swimmers.

For clinical reference to MRI data, femoral neck areal bone mineral density (aBMD) and respective section modulus (Z, an index of bending strength) were determined with dual energy X-ray absorptiometry (DXA, GE Lunar Prodigy Advance, Madison, WI) (29). The aBMD reflects simply the mean thickness of bone mineral within the scanned bone cross-section (or volume) and the Z value the distribution of bone mineral along the scan line of the planar DXA image (10). In our laboratory, the reliability coefficients (R) of repeated in vivo femoral neck aBMD and Z measurements are more than 0.95 and 0.90, respectively.

MRI

MRI imaging was performed using a 1.5T MRI system (Siemens, Avanto, version Syngo MR B15, Erlangen, Germany). First, sagittal, axial, and coronal images of the pelvic region on the side of the dominant leg were obtained with two localization series, and these scout images were then used to specify the appropriate imaging plane orientation for the sequence of the proximal femur, which was done obliquely perpendicular to the femoral neck axis. Body matrix coil was used in combination with three elements of spine matrix coil. Normalization filter was used for the coil profile correction. When surface coils are used the signal from the tissue close to the coil elements is high and decreases rapidly with increasing distance from the coil. Normalization filter can correct the coil sensitivity profile and thus minimize inhomogeneities in image intensity.

The two imaging sequences used in the texture analysis were the following. (i) An axial 3D T1-weighted FLASH (Fast Low Angle SHot) sequence with interpolation in slice selection direction. The acquisition parameters were repetition time (TR) 15.3 ms, echo time (TE) 3.32 ms, slice thickness 1.00 mm without gaps, pixel size 0.91 mm × 0.91 mm, flip angle 10°. (ii) An axial 3D T2* weighted sequence called MEDIC. In MEDIC, multiple echoes, 3 in our sequence, acquired in a gradient echo scan are combined into an image for less artifacts and higher signal to noise ratio (SNR); the early echoes provide increased SNR, whereas later echoes boost contrast. MEDIC is a heavily T2* weighted spoiled gradient echo sequence that uses a series of identically phase encoded gradient echoes, sampled per line in k-space. Unipolar frequency encoding gradients are used to achieve flow compensation and to eliminate resonance effects. For each echo, the magnitude images are reconstructed and postprocessed by using a sum of squares algorithm to improve SNR. The increased receiver bandwidth reduces the T2* effects and impairment of the spatial resolution. The acquisition parameters for MEDIC sequence were TR 40 ms, TE 17 ms, slice thickness 1.00 mm, pixel size 0.91 mm × 0.91 mm, flip angle 12°. In both series, the whole proximal femur was imaged from the femoral caput to the subtrochanteric level of the femoral diaphysis, resulting in 120 image slices. Scan time for FLASH sequence was 5 min and for MEDIC sequence 6 minutes.

In this study, the anatomical level of interest for texture analysis was the femoral neck at the insertion of articulation capsule, which represents a cross-section of the proximal femur that is subject to specific load-bearing in different exercises but is not loaded through direct muscle attachments (29). Furthermore, this region could be distinctly determined from the image series according to clear anatomical landmarks. The image selection for texture analysis was performed manually with a DICOM viewer Osiris (Windows version 4.19, The Digital Imaging Unit of the Service for Medical Computing of the University Hospitals of Geneva, Switzerland).

Texture Analysis

The texture analysis application MaZda (3.20) (23) was used for the automated calculation of texture parameters from selected DICOM MRI images of the femoral neck. For the analysis, the periosteal boundary of the bone cross-section was first segmented, and the cortical region was concentrically peeled off. Then the remaining trabecular region was divided into inferior (I), anterior (A), superior (S), and posterior (P) regions of interest (ROI) (Fig. 1), and these regions were analyzed separately for texture parameters for both imaging sequences (Fig. 2).

Figure 1.

For calculation of texture parameters from the oblique MRI scan of the femoral neck at the level of articulation capsule insertion (left image), the bone cross-section was first segmented (middle), then the cortical bone was concentrically peeled off as indicated by the gray boundary and finally the remaining region was divided into four anatomic sectors (anterior [A], posterior [P], superior [S], and inferior [I]) (right).

Figure 2.

FLASH image (left) and MEDIC image (right). Original images are shown, on the top, zoom-in of regions of interest in the middle, and sector-specific ROIs are drawn on bottom row: inferior (light blue), anterior (green), superior (pink), and posterior (dark blue).

For the texture analysis, the gray level intensity of each ROI was normalized separately by limiting the intensities within the range [μ−3σ, μ+3σ], where μ is the mean gray level value and σ is the standard deviation. Three co-occurrence matrix texture parameters (angular second moment, entropy, and sum entropy) were chosen for this study. The rationale for choosing these parameters is that they have previously shown significant correlation to trabecular structural indices (20). Specifically, the angular second moment and entropy has been shown to correlate with Tb.Sp. and Tb.N, and the sum entropy has been shown to correlate with BV/TV. In the present study, we calculated each co-occurrence parameter for four distances between pixels of interest (d = 1, 2, 3, and 4) and further examined each distance in four directions: horizontal (0°), vertical (90°), 45°, and 135°. The mathematical notations of the parameters are presented in (30).

Statistical Analysis

Statistical analysis of data was performed using SPSS for Windows (version 16.0.2). Due to the skewed distributions of texture parameters, nonparametric statistical approach was chosen for group comparisons. First, Kruskal-Wallis test was used to assess whether there were statistically significant differences between the six study groups in general. If so indicated, the post hoc analysis based on Mann-Whitney test was performed to detect which specific study groups differed from each other. Because there were 15 comparisons in the Kruskal-Wallis test, we considered P < 0.0033 significant enough to indicate group-difference in general. For post hoc tests, a less stringent criterion P < 0.01 was used, because the primary interest was in comparing whether the five exercise loading groups differed from the reference group.

Reproducibility of the texture parameters was evaluated by comparing the results obtained from analyses of two adjacent image slices in a random order. This approach was considered to adequately mimic the situation when a repeated scan is performed and a slightly different image slice may be analyzed. As a measure of reproducibility, the reliability coefficient (R) was calculated as follows: one minus the ratio of the variance (i.e., standard deviation squared) of differences between the texture parameters obtained from the analyses of two adjacent slices to the variance of the given texture parameter in the population. The R-values were calculated for the above-mentioned three texture parameters (angular second moment, entropy, and sum entropy) for both imaging sequences. However, to limit the amount of data, the reproducibility of the texture parameters was determined only for all directions of the very distance that according to statistical analyses most consistently discriminated between the groups. Likewise, the Spearman rank order correlation of these parameters with the femoral neck BMD and Z was determined.

RESULTS

According to Kruskal-Wallis tests, significant (P < 0.0033) between-group differences were indicated in 27 out of 48 different texture analyses (three parameters with four distances and four directions) obtained from FLASH images but in none obtained from MEDIC images. Interestingly, more than 90% of significant between-group differences observed in subsequent post hoc tests were accumulated in the superior sector of the femoral neck, and noteworthy, exclusively in the O-I and H-M groups. Figure 3 illustrates the group-differences in angular second moment, entropy and sum entropy at the distance of 3 pixel and angle 45° in different study groups. This combination indicated the most significant difference in the Kruskal-Wallis test (P = 0.001).

Figure 3.

Boxplot diagrams (median, 25th and 75th quartiles, and 1.5× interquartile range [whiskers]) of the co-occurrence parameters angular second moment, sum entropy and entropy for superior sector of femoral neck for athlete groups (high-impact [H-I]; odd-impact [O-I]; high-magnitude [H-M]; repetitive, low-impact [L-I]; repetitive, nonimpact [N-I]) and nonathlete reference subjects (R) are shown. The texture parameters were calculated at distance = 3 pixels, angle = 45°. The left side diagrams represent the results of the FLASH images and the right side diagrams represent the results of the MEDIC images. The circles represent the outlier values (> 1.5× interquartile range) and the stars represent the extreme values (> 3× interquartile range). Within FLASH images, these parameters indicated the most significant difference in the Kruskal-Wallis test (P = 0.001), whereas MEDIC images did not show statistically significant differences.

Furthermore, the distance of 3 pixels indicated most consistently significant between-group differences in texture parameters (data not shown) and was thus chosen for the reproducibility analyses. Table 1 shows the R-values for the three evaluated texture parameters at the distance of 3 pixels in four anatomic sectors for both imaging sequences. The high R-values were mostly above 0.95 except for the sum entropy which showed somewhat lower values. Overall, the reproducibility of all three texture parameters was good in all four anatomic sectors and appeared also to be independent of the angle or imaging sequence.

Table 1. Reliability Coefficients* of the Three Texture Parameters in each Femoral Neck Anatomic Sector at Distance of 3 for FLASH and MEDIC MRI Sequences
Angular second momentSum entropyEntropy
S(3,0)S(0,3)S(3,3)S(3, -3)S(3,0)S(0,3)S(3,3)S(3, -3)S(3,0)S(0,3)S(3,3)S(3,-3)
  1. *Reliability coefficient is defined as: R = 1 – (SD of differences between the results from adjacent slices)2 / (SD of the given parameter)2 (see text for details).

FLASH/MEDIC          
Anterior           
0.97/0.950.97/0.970.96/0.970.99/0.990.92/0.890.96/0.950.95/0.940.96/0.930.95/0.960.97/0.980.97/0.980.97/0.98
Inferior           
0.98/0.950.98/0.970.94/0.980.99/0.990.93/0.890.94/0.950.94/0.940.96/0.920.95/0.960.97/0.980.96/0.980.97/0.98
Superior           
0.98/0.940.98/0.960.97/0.960.99/0.980.93/0.880.94/0.940.94/0.930.96/0.920.95/0.950.96/0.970.96/0.970.97/0.97
Posterior           
0.98/0.950.98/0.970.95/0.970.99/0.990.93/0.900.94/0.950.95/0.940.96/0.930.96/0.960.97/0.980.96/0.980.97/0.98

Table 2 shows the correlations between texture parameters at the superior sector and DXA-measured femoral neck aBMD and Z. The consistently somewhat greater correlations of FLASH-sequence based texture parameters with aBMD and Z suggest that the FLASH images captured more of the information present in the DXA images than the MEDIC images. It is also interesting to note that the correlations of texture parameters with aBMD and Z were always opposite, whereas the correlation between aBMD and Z is positive and much higher (r = 0.66) than the texture-based weak to moderate correlations (r = 0.12 to 0.41).

Table 2. Correlation* of the Three Texture Parameters at the Superior Femoral Neck With DXA-Measured Areal Bone Mineral Density (aBMD) and Section Modulus (Z) of the Femoral Neck for FLASH and MEDIC MRI Sequences
Angular second momentSum entropyEntropy
S(3,0)S(0,3)S(3,3)S(3, -3)S(3,0)S(0,3)S(3,3)S(3, -3)S(3,0)S(0,3)S(3,3)S(3,-3)
  • *

    Statistically significant (p<0.01) correlations are given in boldface type.

FLASH/MEDIC          
aBMD           
0.40/0.330.34/0.260.41/0.310.33/0.29-0.38/-0.37-0.24/-0.27-0.36/-0.41-0.28/-0.33-0.32/-0.41-0.26/-0.35-0.31/-0.41-0.28/-0.33
Z           
-0.19/-0.17-0.27/-0.23-0.19/-0.17-0.26/-0.200.22/0.120.29/0.250.17/0.120.21/0.190.18/0.170.26/0.230.19/0.170.26/0.22

DISCUSSION

Our study demonstrated that co-occurrence-based texture parameters are able to detect exercise load-associated differences in the trabecular region of the femoral neck. Different trabecular bone texture in female athletes who participated in odd-impact load exercise (soccer and squash players) and high-magnitude load exercise (power-lifters) was indicated. Specifically, the angular second moment, sum entropy and entropy parameters at the superior femoral neck sector were distinct from those among nonathlete reference group.

Our findings are consistent with the fact that bone architecture adapts specifically to loading direction (31). Bone architecture is altered if a new regular loading pattern differs substantially from the previous predominant loading environment and is long-lasting. Odd-impact loading sports represent loading that comprises high accelerations and decelerations (forces) from unusual directions and differ in this respect from other impact sports, which mainly involve maximal vertical loads. Regarding the vertical impacts, the principal loading direction is similar to that which occurs during normal locomotion (i.e., walking and running); obviously, the magnitude and rate of loading is higher. Similarly, power-lifting includes specific squat exercise and deadlift movements that impose the femoral neck to maximal, slow rate muscle forces delivered from unusual directions.

We recently found that odd-impact athletes consistently had 15–20% thicker cortex around the femoral neck compared with nonathletic referents while their femoral neck superior cortex was relatively the thickest among all five exercise loading groups (29). In contrast, high-magnitude loading was not associated with cortical thickness, which was similar to referents. As regards to femoral neck aBMD, the odd-impact group's mean value was approximately 20% higher than in referents whereas the high-magnitude group's mean value was less than 10% and not statistically different from the referents. Nevertheless, it is possible that both of these exercise loadings can specifically affect trabecular bone structure within the femoral neck superior region, as suggested by the present findings of texture parameters. Due to the limited resolution of most in vivo imaging techniques, including the present study, and the scarcity of relevant studies, the influence of different exercise loadings on actual trabecular architecture in human bones (i.e., thickness, number, separation, and orientation of trabecular elements) is not yet well known (32). The degree of mechanical loading due to type of physical activity has been associated with trabecular microstructure in young men in a large study based on high-resolution, three-dimensional, peripheral quantitative computed tomography at the distal tibia and radius (33). Also, two small studies using high-resolution MRI have found that the trabecular number was increased in the knee region both among gymnasts and Olympic fencers, but no information was provided about spatial distribution of trabeculae (26, 27). To our knowledge, the present study provides novel information on the association of exercise loading with trabecular texture at the clinically relevant femoral neck.

Of interest, odd-impact and high-magnitude exercise loading seemed to specifically influence the superior sector of the femoral neck, which may have some clinical bearing. In a recent study comparing QCT-measured sector-specific trabecular bone mineral density with hip fragility (2), trabecular density was found to decline asymmetrically with aging and the loss was most rapid at the supero-posterior region of femoral neck. This seminal finding by Thomas et al (2) highlighted not only the importance of the superior femoral neck region on bone strength but also that of comprehensive assessment of the whole-bone structure, not just a single compartment (10). It remains to be investigated how intensive and long the specific physical loading should be to translate the exercise related changes both in cortical and trabecular bone structure into reduced fracture risk in old age.

In the present study, the trabecular bone texture was assessed with three co-occurrence matrix-based second order parameters, which describe the spatial relationship between pixel pairs in four directions with four pixel distances. While statistically significant group differences were indicated from short to long offsets of pixels, the distance of three pixels most consistently provided significant differences between the study groups. Of the two MR imaging sequences used in this study, the FLASH images provided more clear consistent results than the MEDIC images, which did not indicate any between-group difference under the stringent criterion for statistical significance. Evidently, this difference cannot be explained by reproducibility, which was similar for both sequences. In fact, the reproducibility of texture analyses appeared to be quite good and fully comparable to commonly used DXA and pQCT methods in bone research (34, 35). Rather, the observed differences in performance may be due to the fact that the tissue characterization results are dependent on the signal intensity rising from the structures imaged. It is thus very likely that the observed texture results were modulated by this factor which is different for T1 FLASH and T2* MEDIC sequences. The fat suppression properties of these sequences differ: the FLASH presents the bone marrow with higher signal than the MEDIC—giving the computerized method better basis to find subtle gray level changes originating from trabelular bone signal. Altogether, the present texture analysis of 1.5T MR FLASH images has potential in evaluating apparent trabecular bone structure and adaptation to exercise loading in different age and target groups without ionizing radiation exposure.

The strengths of the present study are the large sample of athletes representing distinct exercise loadings and the comparison of two different established 1.5T MRI sequences for image acquisition. Despite recent advances in magnetic field strengths, 3T or higher field scanners are not yet widely used in most clinical imaging centers. Therefore, we specifically concentrated on imaging protocols available for clinical purposes in the majority of centers and the 1.5T scanner was used. The major weakness of our clinical imaging protocol is the spatial resolution of 0.9 mm, which did not allow true viewing of single 0.1- to 0.15-mm-thick trabeculae. Instead, the trabecular bone appearance was estimated by means of gray level co-occurrences between pixels. Limited spatial resolution is a known disadvantage in the clinical MR imaging environment. Therefore, increased imaging resolution might lead to even better separation of athletes from referents and reveal smaller textural details. Also, the obvious disadvantage of using textural features compared with the use of direct measures of trabecular bone architecture by high-resolution imaging is the impossibility to interpret the findings in tangible terms of bone structure. However, the textural parameters appear to reflect the underlying trabecular architecture (13, 20), but further validation is needed to verify the utility of the present approach.

As a clinical application, the manual phases of the present textural approach are critical for the quantitative image analysis, the successful execution of which requires trained personnel. Accurate location and alignment of the image plane with scout images is essential for image analysis to be performed from certain directions or angles in respect to anatomical structures or landmarks. Selection of the image slices on the basis of appropriate anatomical landmarks from large image stacks is cumbersome and requires expertise in anatomy and computer applications. The manual or semi-automatic segmentation of ROIs by drawing the areas on images with a texture analysis application seems to be the most time consuming and pedantic manual part of the analysis procedure. While the texture analysis application MaZda proved stable and reasonable to use for these phases, our study setting with six exercise loading groups and division of the trabecular bone compartment into four anatomic sectors was challenging and needed some manual analysis despite the use of the software package. Obviously, for analyses of greater quantities of materials and for future clinical use, the texture analysis procedures need to be automated and optimized to minimize the time required for specialists to perform the analyses.

Recently, promising results from three-dimensional (3D) texture analyses have been reported. For example, comparison of 2D and 3D co-occurrence matrix parameters in discrimination of cerebral tissue, tumor, necrosis and edema indicated superior classification with the 3D data compared with the 2D data (36). Also, 3D texture analyses of intracranial tumors based on co-occurrence and run-length matrices were shown to outperform the 2D analyses in discriminating primary tumors from metastatic tumors, whereas benign tumors were similarly discriminated from malignant tumors with both 2D and 3D approaches (37). In our study, the 2D approach was chosen primarily because 2D is yet the most common and familiar presentation of images in clinical centers; radiologists select the image slices and interpret the images in 2D. Thus, drawing ROIs and performing texture parameter calculation in 2D would be less cumbersome and require less time than the 3D analysis. However, the trabecular bone is fundamentally a 3D structure, and application of texture analysis in 3D might enhance the structural analysis in comparison to 2D analyses. Obviously, because of limited resolution of MRI in clinical settings, the ability of 3D analysis to reveal clinically significant information not captured by 2D analysis remains yet to be investigated.

In conclusion, MRI texture analysis provides a quantitative method for detecting and classifying apparent structural differences in trabecular bone, which are associated with specific exercise loading patterns.

Acknowledgements

L.C.V.H. is funded by research grants from the Finnish Medical Society Duodecim and Medical Fund of the Pirkanmaa Hospital District. R.N. is funded by research grants from the Academy of Finland, Finnish Cultural Foundation, and Medical Fund of the Pirkanmaa Hospital District.

Ancillary