Introduction

With recent advances and improved availability of ultra-short echo-time (UTE) imaging sequences¹, musculoskeletal MRI is increasingly focusing on imaging of tendons. In particular, the quantification of relaxation parameters holds a high potential for the characterization of various diseases, such as tendinopathy, as well as the general condition of tendons^2,3. One major challenge, however, is the segmentation and visualization of tendons, because they can be small compared to other tissues and can have curved structures. Segmentation of tendons is made even more difficult by the fact that UTE images typically show very little contrast between tissues due to the very short echo-times. In this work, we present an automated segmentation approach of the patellar and quadriceps tendons based on bivariate histograms of volumetrically measured T₂^* and T₁ relaxation times.

Methods

Data acquisition was performed using a 3D radial center-out UTE sequence with short hard-pulse excitation¹. For T₂^* estimation, corresponding power images from a monopolar multi-echo readout with echo-times (TE) of 0.10ms, 2.48ms, and 4.90ms were fitted voxel-wise using a squared exponential function that included an additional offset parameter to account for potential noise bias⁴. T₁ relaxation times were estimated from a variable flip angle acquisition using five flip angles of 5°,12°,20°,30°, and 38° and by performing a two-parameter fit to the fast low-angle shot (FLASH) gradient-echo signal equation3. Other acquisition parameters were: 80×61×50 matrix, (160×123×100)mm³ field of view, (2.0×2.0×2.0)mm³ isotropic spatial resolution, 20ms repetition-time and a 125kHz readout bandwidth. Measurements were performed on five volunteers, aged between 24 and 50 years old, without known pathologies on a 3T whole-body MRI scanner (Magnetom PRISMA, Siemens Healthineers) using a 16-channel NORAS Variety flex coil (NORAS MRI products GmbH). Images were reconstructed offline with MATLAB using re-gridding with iterative sampling density compensation and an optimized kernel⁵. Prior to further processing, the reconstructed 3D relaxation parameter maps were masked to exclude contributions from noise regions outside the knee. To calculate the bivariate histograms, the T₁ and T₂^* relaxation times of all voxels were binned into equally sized bins. Clearly visible clusters in the histograms were manually outlined by drawing ROIs around them, and subsequently visualized using 3D surface reconstruction. To improve the masks resulting from this histogram analysis and to remove outliers, a connectivity analysis was performed prior to visualization, retaining only the largest connected components of the masks. Since no prominent cluster was found in the bivariate histogram located in the range of the expected T₁ and T₂^* times of the patellar and quadriceps tendons, a ROI was placed in the histogram encompassing a rectangular region in the ranges of 1.0ms<T₂^*<3.0ms and 350ms<T₁<900ms.

Results

Maps of the T₂^* and T₁ relaxation parameters are shown in Figure 1 for a single subject, demonstrating that both the patellar and quadriceps tendons were clearly identifiable. Since the longest echo-time was only 4.9 ms, the values obtained for other tissues with longer T₂^* are more uncertain and most likely underestimated. The bivariate histogram (Figure 2) revealed several well-delineated clusters, exhibiting similar relaxation parameters in the range of 150ms<T₁<450ms and 3ms<T₂^*<7ms. Also visible is a broader cluster in the range of 800ms<T1<1500ms and 5ms<T₂^*<11ms. By separating and masking the bivariate histogram, the tissue types underlying the clusters became evident as shown in Figure 3. The narrow and sharp clusters between 150ms< T₁<450ms and 3ms<T₂^*< 7ms mainly reflect bone marrow (Figure 3, blue) as well as fat and skin (Figure 3, green), while the broader cluster with higher T₁ values (Figure 3, yellow) can be attributed to muscle tissue. By placing an ROI in the range of 1.0ms<T₂^*<3.0ms and 350ms<T₁<900ms in the bivariate histogram, the patellar and quadriceps tendons could also be extracted (Figure 3, red). Based on the segmentation of both tendons (Figure 3, red), a color-coded 3D volumetric rendering was created to visualize the distributions of T₁ and T2* values over the entire tendon volume (Figure 4), including parts of the entheses. Inspection of these volumetric renderings indicates that T₁ is not constant over the volume of the tendons, but contains “hot spots” in some subjects as well as an increase in the relaxation times towards the entheses.

Discussion and Conclusion

Using bivariate histograms of T₂^* and T₁ for 3D segmentation of the quadriceps and patellar tendons allowed for a fully automated approach and required only one-time user interaction to define the relaxation time bins used for segmentation of the tendon voxels. Visualizing volumetric relaxation maps holds the potential for future studies to assess these parameters not only over manually selected region-of-interests, but across the entire tendon volume. As demonstrated (Figure 4), T₁ within the tendons varies individually, implying that placement and size of any region-of-interest based analysis highly affects mean values. Our results suggest that a bivariate histogram analysis can also be applied to other tissues types, such as fat or muscles. For this purpose, however, acquisition of data with longer echo-times would be recommendable to avoid underestimation of T₂^* and compression of the T₂^* bins along the x-axis of the histograms. Additionally, further optimized and accelerated^6,7 protocols for faster T₁ mapping would help to reduce the relatively long scan time of currently 22 minutes or alternatively to increase spatial resolution

Acknowledgements

No acknowledgement found.