Automatic hip abductor muscle fat fraction estimation and association with early OA cartilage degeneration biomarkers

1 INTRODUCTION

Osteoarthritis (OA) affects over 30 million American adults, primarily affecting the weight-bearing joints such as the knee and hip.¹ While the main sign of OA is cartilage degeneration, it has been recognized as a whole joint disease, and includes pathological changes like progressive proteoglycan loss and collagen disruption, thickening of the subchondral bone, osteophyte formation, bone deformation, and muscle degeneration.^{2, 3} Femoroacetabular impingement (FAI) is the symptomatic abutment between the proximal femur and the acetabular rim, and there is evidence that supports the link between FAI and hip OA.^{4, 5} To support this, abnormal hip morphology (such as cam impingement due to a nonspherical head or excessive coverage of the femur head by the acetabulum) has been identified as an important factor in the development of hip OA.^{6, 7} Studies have reported that 90% of patients with primary hip OA had an abnormal bony morphology.⁸ FAI is a pathomechanical process that may lead to early OA in young and active adults, making FAI patients a good cohort to model early development of hip OA.

While it is known that changes in bone morphology can lead to OA, the relationship between bone and muscle is important to understand and develop noninvasive interventions that can target muscle growth, leading to bone growth, slowing down OA progression. It has been proposed that blood flow to limbs is proportional to muscle mass, leading to bone growth.⁹ Similar genetic factors influence bone and muscle, and a close relationship between bone and muscle is observed via the Indian hedgehog pathway and fibroblast growth factor-2.⁹ It has also been shown that muscle cells may play an important role in regulating cartilage gene expression.¹⁰

Thus, in an effort to recognize the effect of OA on muscles, muscle atrophy, and weakness around the hip have been studied and identified in hip OA.¹¹ For the hip, three abductor muscles act as joint stabilizers in weight-bearing conditions: the gluteus medius (GMed), gluteus minimus (GMin), and tensor fasciae latae (TFL).¹² A study has shown that fatty inclusions in muscle reduce muscle force and quality.¹³ Fatty infiltration in muscle has been compared with the effects of neurological changes, hormonal changes, and inflammatory cytoknines as a potential contributor to loss of muscle force production.¹⁴ Intramuscular fat has shown to potentially have effects on insulin resistance, inflammation, and fracture risk among others.¹⁵ Type I and II muscle fiber atrophy have been correlated to the length of disease duration in OA.¹⁶ A previous study has found higher intramuscular fat fractions in the quadriceps in people with knee OA.¹⁷ Increased intramuscular fat content has been identified as a risk factor for OA due to its effect on metabolic pathways.¹¹ Additionally, muscle power has been found to be an independent determinant of pain and quality of life in knee OA.^{18, 19} A previous study showed that in addition to hip OA, patients with FAI also experience a functional disability during dynamic weight-bearing activities, and have been found to have muscle weakness in their hip, especially a reduced ability to activate the TFL during flexion.²⁰ A previous study, using a small sample size, found that patients with late-stage OA with cartilage degeneration have reduced volumes of the GMed and GMin and increased fatty infiltration of the GMin.²¹ Given the evidence of quadriceps muscle weakness in knee OA, quadriceps strengthening has been suggested for the management of knee OA.¹¹ However, it remains unclear whether muscle weakness, as seen in knee OA, is observed in hip OA, and therefore, a target for muscle strengthening for hip OA needs to be identified.²²

OA progression is often measured using joint space narrowing as quantified by radiograph-based Kellgren-Lawrence (KL) scores, minimal joint space width, Croft scores.^23-25 However, compositional change in articular cartilage such as proteoglycan depletion and water loss may occur before joint space narrowing is seen on radiographs. T_1ρ, T₂ relaxation times provide a measure related to proteoglycan content and collagen orientation, respectively, reflecting the cartilage biochemistry and composition.^{26, 27} In a previous longitudinal study, patients with hip OA, who demonstrated elongated T_1ρ, T₂ relaxation times, developed morphological cartilage degeneration, indicating that T_1ρ, T₂ relaxation times may be biomarkers for early OA.²⁸ Due to the abnormal hip joint contact, elevated T_1ρ and T₂ relaxation times, similar to as seen in OA patients have also been observed in FAI patients.^{29, 30}

OA is a whole joint disease that affects bone, cartilage, and muscle, with no consensus on which changes occur first. There are no clear links between the fat content in muscle and early hip OA characteristics like cartilage lesions and are furthermore hampered by the fact that manual muscle segmentation and processing make it difficult to accurately extract metrics characterizing muscle. Thus, there is a need to determine if muscle degeneration (reduction in volume and/or increase in fat content) occurs in early OA and if it is associated with cartilage degeneration, to diagnose OA early and potentially slow down disease progression by finding targets for intervention.

To fill these gaps, and address the hypothesis that hip abductor muscle fat is different between patients who have early OA, advance OA and FAI, the goal of this study is twofold: (i) to develop an automatic segmentation and quantification pipeline to estimate the fat fraction and volume in the GMed, GMin, and TFL muscles (ii) to study associations of hip abductor muscle fat fractions with patient age, patient gender, T_1ρ and T₂ relaxation times.

2 METHOD

2.4 Model performance evaluation and muscle fat and volume quantification

Performance of the automatic segmentation was evaluated using Dice coefficient which was computed using the manually segmented masks and those predicted by the V-Net model. Details on the Dice coefficient are provided in Appendix B. While commonly used, Dice has a shortcoming where it does not consider the positions of voxels that are not in the overlap region, which means that they provide the same value independent of the distance of the voxels.³⁵ Additionally, Dice does not give information on the accuracy of the edges of the segmentation, and is automatically greater in larger areas. Thus, to use another efficiency metric besides Dice similarity coefficient, and to visualize where the differences between the manually and automatically computed masks are located, a distance map for each patient was modeled by using the displacement of the vertices in each mask.^{36, 37} The mean distance (in mm) was computed for each patient using the resolution and the number of pixels seen in the maps.

The fat fraction in each muscle was calculated as shown in Equation 1 ³⁸; by using the number of pixels in the segmented region in the fat only and water only images obtained using the IDEAL sequence.

(1)

The fat fractions and volumes of each muscle using the predicted and manually segmented masks were calculated and compared to check for interchangeability between manual and automatic segmentation, by plotting a correlation and Bland-Altman plot to compare the two methods of segmentation, and extracting the Pearson correlation coefficient values between the manual and automatic segmentation. Additionally, a Student's t test was used to test if there were any significant differences in the extraction of fat fraction between the two V-Net models. The significance level was set at 0.05.

Figure 1 shows the automatic segmentation and quantification pipeline built and used in this study.

3 RESULTS

3.1 Performance of the automatic segmentation pipeline

For the manually segmented masks used for training the V-Net, the ICC was computed as 0.94, which ensured that there was minimal variability in the segmentations between the two observers. An example of the manual and automatic segmentation in three slices of each muscle for each V-Net can be seen in Figure 2. The first V-Net provided a mean Dice coefficient of 0.90, 0.88, and 0.91 for the GMed, GMin, and TFL, respectively, on the test set. The Bland–Altman plots showing the difference in fat fractions and volumes in each of the muscles on the test set using the manual masks and the V-Net generated masks are shown in Figure 3. The mean distances computed with the distance mapping were 1.15, 1.19, and 0.74 mm on the GMed, GMin, and TFL, respectively, on the same test set. A mean distance map for each of the muscles is seen in Figure 4. Performance metrics of the ensemble V-Net can be seen in Appendix B. For each patient volume, the algorithm took ~2 s to predict all three muscle segmentations. Going forward, the first V-Net was used to derive the associations, since it had better performance in terms of Dice scores and muscle fat and volume estimation. Additionally, the lack of significant difference between the performance of the two models, tested by the difference in fat fractions computed by each model (p > 0.1), justified the use of a less complex model for the segmentation.

3.3 VBR results

The results for the voxel-based derived associations between muscle fat fractions with T_1ρ relaxation times in the whole patient and stratified (KL = 0,1, KL = 2,3, FAI) cohorts can be seen in Figure 5. The results for the voxel-based derived associations between muscle fat fractions with T₂ relaxation times in the whole patient and stratified cohorts can be seen in Table 2.

Table 2. Summary of associations between cartilage T₂ relaxation times and muscle fat

		All	KL = 0, 1	FAI	KL = 2, 3
Tensor fascia latae	PSV	19.50%	2.05%	27.47%	5.35%
	R	−0.32	0.54	−0.55	−0.71
	P	0.02	0.02	0.02	0.02
Gluteus medius	PSV	18.94%	5.55%	10.21%	5.40%
	R	−0.33	−0.52	−0.55	−0.72
	P	0.02	0.02	0.02	0.02
Gluteus minimus	PSV	29.42%	8.07%	30.26%	7.59%
	R	−0.33	0.52	−0.55	−0.71
	P	0.02	0.02	0.02	0.02

• Abbreviations: PSV, percentage of significant voxels; R, mean partial Pearson correlation value; P, mean p value.

Table A1. Demographics and Clinical Characteristics (HOOS survey outcomes)

	OA cohort	FAI cohort
Male	23 (60%)	14 (60%)
Female	15 (40%)	9 (40%)
Age	53.39 ± 12.62	33.95 ± 9.81
BMI	23.65 ± 2.55	24.00 ± 3.44
Pain	91.77 ± 14.03	62.27 ± 18.45
Symptom	90.39 ± 12.86	60.90 ± 15.70
Quality of life	87.17 ± 17.34	29.26 ± 14.22
Sports/recreation	92.26 ± 14.34	39.48 ± 19.42
Function in daily living	94.69 ± 12.34	67.31 ± 20.79

• Abbreviations: BMI, body mass index; HOOS, Hip Disability and Osteoarthritis Outcome Score.

As seen in the Figure 5, for the TFL, in the whole patient cohort, 6.81% of voxels have a weak negative correlation (R = −0.31, p = 0.02) with the fat content. However, in the stratified results, it can be seen that 8.32% of the voxels have a moderate negative correlation (R = −0.53, p = 0.02) with the fat content in the FAI cohort. It can be seen that in the approximate same voxel locations in the cartilage, the advanced OA patients exhibit a strong negative correlation (R = −0.75, p = 0.02) in 4.78% of the voxels.

Similarly, as seen in Figure 5, for the GMed, in the whole patient cohort, 6.13% of voxels have a weak positive correlation (R = 0.33, p = 0.02) with the fat content. However, in the stratified results, it can be seen that 7.42% of the voxels have a moderate positive correlation (R = 0.53, p = 0.02) with the fat content in the FAI cohort. It can be seen that in the approximate same voxel locations in the posterior cartilage, the advanced OA patients exhibit a strong positive correlation (R = 0.71, p = 0.02) in 3.42% of the voxels. The OA and FAI patients also have the similar location of moderate positive correlations in the anterior femoral cartilage.

For the GMin, as seen in Figure 5, in the whole patient cohort, 21.5% of voxels have a weak negative correlation (R = −0.37, p = 0.02) with the fat content. In the stratified results, it can be seen that 16.1% of the voxels have a moderate negative correlation (R = −0.53, p = 0.02) with the fat content in the FAI cohort. It can be seen that in the approximate same voxel locations in the posterior cartilage, the advanced OA patients exhibit a strong negative correlation (R = −0.73, p = 0.02) in 6.31% of the voxels. The OA and FAI patients also have similar location of moderate positive correlations in the weight-bearing region of the acetabular cartilage.

Similar results for T₂ relaxation times associations with fat content are seen in Table 2. For TFL, the FAI cohort shows a moderately negative correlation (R = −0.55, p = 0.02) in 27.47% of voxels with fat content, whereas in the whole patient cohort 19.50% of voxels show a weak negative correlation (R = −0.32, p = 0.02). Similarly, in the GMed, the correlation is stronger in the FAI cohort (R = −0.55, p = 0.02, PSV = 10.21%) than in the whole cohort (R = −0.33, p = 0.02, PSV = 18.94) and the late stage OA has strong negative correlations (R = −0.72, p = 0.02) in 5.40 of the voxels. The most voxels with correlations are seen in the GMin, where again, the correlation is stronger in the FAI cohort (R = −0.55, p = 0.02, PSV = 30.26%) than in the whole cohort (R = −0.33, p = 0.02, PSV = 29.42) and the late stage OA has strong negative correlations (R = −0.71, p = 0.02) in 7.59 of the voxels.

4 DISCUSSION

In this study, we showed a feasible, efficient, and fully automatic pipeline to segment the hip abductor muscles and compute their fat fraction and volumes. We observed associations between the fat fractions of the hip abductor muscles and early markers of cartilage degeneration, T_1ρ and T₂ relaxation times, in patients with hip OA and FAI.

Muscle volume and fat fraction quantification have previously been carried out by semi-manual to fully manual boundary tracing techniques that can be feasible, but are often time consuming and prone to human error. Additionally, manually retrieving the inter-muscular fat fraction can be more challenging due to the blurry boundaries between muscles when fat is present between different muscles.⁴¹ Therefore, automated muscle segmentation has been an area of active investigation. Other techniques proposed toward this goal include fuzzy clustering, manual seeding, intensity scaling, and registration-based segmentation.^{41, 42} While all of these have produced acceptable results, they mostly were applied to 2D images with a limited number of slices that may not cover the entire muscle. The fully automated segmentation and fat fraction and volume quantification pipeline developed by this study was applied to 3D datasets so that it can maximize the coverage of the entire muscle. Additionally, this was the first fully automatic, 3D MR image muscle segmentation pipeline developed for the hip muscles, allowing a greater coverage than a single slice segmentation, thus providing a more complete estimation of fat fraction and volume estimates as compared to those obtained in single slices. Applications of this pipeline could be extended to evaluate muscles around the knee and other joints, for OA, measuring muscle strength, degeneration, evaluation of effects of exercise, and other musculoskeletal studies.

We expected the ensemble V-Net to perform better since it was given more contrast types as inputs and was narrowed down to one muscle only, however, in our case the simpler V-Net showed a slightly better performance. While we expected the additional information to boost the V-Net performance,⁴³ in this case, it was possible that since only 51 patients were used to train the V-Net, there was no utility of the additional information. A larger number of patients would be useful to use the ensemble V-Net where it would have shown a better performance by being invariant to more variability in the data set.

We observed difference in mean muscle volumes between men and women, with women having smaller GMed and GMin volumes as compared to men, which has been previously observed in overall muscle differences between men and women.⁴⁴ This effect could be due to overall size (women are smaller than men in general), and muscles of the lower limb have been shown to scale with height and body mass.⁴⁵ Our study found that, in general, the TFL muscle had negative associations between fat content and T_1ρ and T₂ relaxation times. Casartelli et al have previously shown that patients with FAI have a reduced ability to activate the TFL muscle, as measured by electromyographic activity, during hip flexion.²⁰ This reduction of ability to activate the TFL could explain the changes in load caused to the anterior portion of the cartilage in patients with FAI, which could also lead to OA. Our study indicated that the GMed muscle had negative associations between fat content and T₂ relaxation times, but positive associations between fat content and T_1ρ relaxation times. These results are important to highlight the different information brought to the table by T_1ρ and T₂ relaxation times, indicating that the GMed fat content has a different association with the proteoglycan content and collagen orientation. A previous study has found that patients with hip OA demonstrated increased muscle activation in the GMed during stepping tasks, most likely a compensatory mechanism to muscle weakness,⁴⁶ which could also explain the positive association seen between the fat content and T₂ relaxation times. The opposing results seen between GMed and GMin could also indicate a complicated compensatory mechanism, as often seen in patients that have pain and reduced ability to symmetrically use their muscles. Using an automatic segmentation pipeline as developed in this study, could further investigate this asymmetric mechanism in a larger patient cohort.

The most interesting results are seen with the GMin muscle, which contains the highest number of voxels in both T_1ρ and T₂ relaxation times associations with the fat content. A previous study has shown increased fatty infiltration in GMin with increasing OA severity, however, our results indicate that the increase in fat in GMin was associated with a reduction T_1ρ and T₂ relaxation times.²¹ There is preliminary evidence that suggest muscle atrophy starts to occur in more advanced stages of OA.⁴⁷ In our advanced OA cohort, we still observed that the muscle fat was associated with a reduction in T_1ρ and T₂ relaxation times. However, it was interesting to see that the location of associations were in the same, anterior regions of the cartilage, indicating that that is the part of the cartilage that may be affected by the GMin fat content. The function of the GMin muscle has also previously been identified as stabilizing the head of the femur in the acetabulum by tightening the capsule.⁴⁸ Since FAI affects the anatomy of the proximal femur head relative to the acetabulum, we expected to see changes in the GMin fat fraction for FAI patients. We observed this effect, where the FAI patients had stronger correlations with cartilage degeneration markers as compared with the whole cohort. A previous study has found a more pronounced burst of muscle activity in patients with increasing severity of OA in the GMin muscle but nothing significant in the GMed.⁴⁹ Putting these results together with those from our study, increased GMin activity, lower fat content of GMin and higher fat content of GMed are associated with OA biomarkers, could maybe indicate a compensatory mechanism that may be occurring early in OA or in FAI patients. However, more data would be needed to support this mechanism and confirm any causal effect.

There are some limitations in this study that need to be addressed. First, the manual segmentations had masks for the main coverage of each muscle, which was about six slices each, as a result of which the V-Net also segmented about six slices of each muscle. While this covers most of the entire muscle, there are some edge slices that are not segmented. Second, our associations were observed on a cohort of 61 patients. A larger cohort would be needed for stronger, more robust correlations and generalizability. While using the FAI cohort to model early OA did give insights into the fat fraction changes in the hip abductor muscles before OA onset, a longitudinal study following the same patient cohort would be useful in determining the progression of events in leading to OA. Within our test patient cohort, the maximum fat fraction calculated was 15%, and in a future study, we will evaluate the robustness of the segmentation model on patients with higher fat fractions upon recruitment. Additionally, the effect of patient activity level (which could have significant effect on muscle size) was beyond the scope of this study, and will be considered in future studies. In this study, we used T_1ρ relaxation times as a biomarker for OA. While there is evidence that shows a strong link between T_1ρ relaxation times and OA linking the biochemical composition of cartilage with T_1ρ, its relationship with proteoglycan is not fully understood.⁵⁰ However, there is research that suggests that T_1ρ has a relationship with OA status,⁵¹ is a predictor of future morphological progression of OA⁵² and has a relationship with patient reported outcomes at a voxel based level,²⁸ which justifies its use in this study.

Nevertheless, using state-of-the-art 3D encoder-decoder convolutional neural networks, we were able to produce fast, accurate, and precise automatic segmentations of the hip abductor muscles that are invariant across patients with OA. Additionally, our model demonstrated efficacy in extracting muscle volumes and fat fractions that can be used in the prediction and monitoring of pathology in OA and other studies. This demonstrates our model's interchangeability with manual segmentation, representing an important step toward the clinical translation of quantitative MR imaging techniques. The associations derived between OA biomarkers and fat fractions in the hip abductor muscles provide insight into the early changes that occur before OA is fully developed, indicating that the GMed and GMin could be targeted for early intervention for OA patients.

CONFLICT OF INTERESTS

The authors declare that there are no conflict of interests.

AUTHOR CONTRIBUTIONS

Study design: Radhika Tibrewala, Valentina Pedoia, Richard B. Souza, and Sharmila Majumdar. Data processing: Radhika Tibrewala. Clinical expertize: Alan L. Zhang and Thomas M. Link. Manuscript draft: Radhika Tibrewala. Statistical expertize: Jinhee Lee. Obtaining funding: Richard B. Souza and Sharmila Majumdar. All authors have approved the final manuscript.