A retrospective evaluation of individual thigh muscle volume disparities based on hip fracture types in followed-up patients: an AI-based segmentation approach using UNETR

Introduction

Hip fractures represent a significant health concern, particularly among the older population. These fractures often lead to a substantial loss of muscle mass and strength, which can contribute to functional decline and increased disability (Yoo et al., 2018; Turkmen & Ozcan, 2019; Groenendijk et al., 2020). Efficient and targeted rehabilitation strategies are, therefore, crucial to mitigate the loss of thigh muscles, which play a significant role in mobility following a hip fracture (Pham et al., 2017; Min et al., 2021; Yoo et al., 2022).

In the realm of hip fractures, several prominent types can be identified, including femoral neck fractures (FNF), intertrochanteric fractures (ITF), greater trochanteric fractures and lesser trochanter fractures. A femoral neck fracture, which takes place in the region connecting the femoral shaft to the femoral head, leads to atrophy in the muscles of the hip flexors, adductors and the gluteal region muscles (Chang et al., 2023). An intertrochanteric fracture occurs between the greater and lesser trochanter, leads to muscle loss in the hip adductor and quadriceps muscle groups (Satone et al., 2022). These insights inform the necessity of targeting specific muscles for rehabilitation post-hip fracture surgery, tailored to the type of fracture sustained. However, quantifying the progress of this specialized treatment can prove to be a challenging task.

The cross-sectional area (CSA) of muscles has been utilized to assess muscle size, intending to identify individuals at risk of sarcopenia and quantify deformities of muscles, the degenerative loss of skeletal muscle mass and strength. However, this approach has several limitations. The primary limitation is that by quantifying the muscle area in a single plane, CSA overlooks the complexity of individual muscle groups (Honkanen et al., 2019). It fails to capture essential aspects such as muscle composition, distribution and overall performance. Factors such as participant positioning, limb orientation and the choice of imaging plane can also influence the accuracy and comparability of CSA measurement.

The volume of each individual muscle, obtained by calculating the annotated segmentation mask, can overcome these limitations (Hiasa et al., 2019). However, manual segmentation on CT scans is a time-intensive, laborious and costly task that requires significant effort and expertise. The process often exhibits high variation due to the difficulty of differentiating tissue characteristics. While CT scans provide excellent visualization of bony structures and dense tissue, they have limitations in differentiating soft tissues such as muscles. Muscles have similar radiodensity, making it a challenge to distinguish individual muscles based on CT scans. Yet, achieving precise voxel-level segmentation is critical for accurately quantifying each individual muscles’ volume and gaining comprehensive insights into muscle performance.

To address these challenges, our previous study proposed a deep learning based automatic individual thigh muscle segmentation approach using the UNETR architecture (Hatamizadeh et al., 2021). Our model demonstrated a high degree of accuracy and precision, achieving a dice score of 0.84 and a relative absolute volume difference of 0.019 $\pm$ 0.017%. The dice score, a statistical measure of similarity, quantifies the overlap between the model’s segmentation output and the ground truth, while the relative absolute volume difference assesses the absolute difference in volume between the segmented output and the ground truth volume, presented as a percentage. This approach leverages the power of deep learning algorithms to learn intricate muscle features and perform precise segmentation at the voxel level. By automating the segmentation process, our proposed method enables efficient and accurate calculation of each individual muscle’s volume (Kim et al., 2024).

The application of our UNETR based model holds promise in advancing muscle volume assessment and enhancing the rehabilitation process by providing valuable insights into muscle performance. Furthermore, our findings provide the groundwork for future research exploring the potential of automatic muscle segmentation models in large cohorts and assessing functional outcomes based on muscle volume changes.

The primary objective of our study is to specifically examine how individual thigh muscle volumes vary in response to different types of hip fractures, focusing on femoral neck fractures and intertrochanteric fractures. By analyzing pre-operative and post-operative CT scans of patients with these fractures, we aim to identify distinct patterns of muscle volume changes. This detailed examination is designed to provide crucial information that can be used to optimize rehabilitation methods for individuals based on hip fracture type.

Materials and Methods

CT scans acquisition

Our investigation included 18 participants, drawn from a cohort of 478 individuals who had been identified with hip fractures. These participants had an average age of 77.3 years old, with a standard deviation of 9.73. Specifically, the average age for females was 78, while for males it was 72.7. Both pre-operative and post-operative CT scans were conducted for each participant in the study. These CT scans were carried out with patients in a supine position (both Head-First Supine and Feet-First Supine), encompassing the entire thigh area from the hip to the knee joint, referred as whole thigh level CT scans. The participants were selected from Gyeongsang National University Hospital during the period from December 2016 to June 2022.

As illustrated in Fig. 1, the criteria for inclusion targeted those with femoral neck fracture and intertrochanteric fracture who were in stable condition, whether to follow up, required evaluation for the skeletal muscle index (SMI) variable and no objections to CT scanning. To uphold the credibility of muscle segmentation, we set exclusion criteria, such as lower limb amputation, femur shaft fracture, subtrochanteric fracture, noticeable muscle or bone deformities and presence of significant artifacts in imaging.

The rationale behind excluding lower limb amputation was the potential for major anatomical variations that might hinder accurate muscle segmentation. Participants who had femur shaft and subtrochanteric fractures were also omitted to prevent potential distortion in muscle appearance due to the angulation of the two portions of the fractured region. This exclusion helped to mitigate any confounding variables associated with the fractured femur.

Additionally, mobility after hip fracture surgery is a crucial factor in the evaluation of patients (Huo et al., 2015). In our study, we assessed mobility using Koval’s grade, focusing on a range from 1 to 6, which present varying degrees of walking ability, from the full scale of 1 to 7. Koval’s grade categorizes walking dependency into three main groups: (1–3) Community ambulatory, (4–6) Household ambulatory, (7) Nonfunctional ambulatory. Our study’s data shows that out of 18 participants, 15 were identified as Community ambulatory (Koval’s grades 1 to 3), while three were classified as Household ambulatory (Koval’s grades 4 to 6).

Through the careful application of these selection criteria, our intention was to establish a consistent and comparable study group, thereby reducing any potential variables that might otherwise influence the analysis of muscle segmentation.

Segmentation labels

In our study, we performed a classification encompassing 30 different classes, including the iliac, femur and background elements, all within five principal thigh muscle groups from hip to knee joint CT scans, referred to as whole thigh level CT scans. These five prominent thigh muscle groups were divided into the anterior, medial and posterior thigh muscles along with the gluteal region muscle and other miscellaneous categories.

Within the anterior thigh, we classified five muscles, namely the sartorius, rectus femoris and vastus muscles, which are further subdivided into the lateralis, intermedius and medialis. In the medial thigh region, five muscles were categorized, consisting of the adductor muscles (magnus, brevis and longus), gracilis and pectineus. The posterior thigh muscles were distinguished as semitendinosus, semimembranosus and biceps femoris. In the gluteal region, eight muscles were classified, including the gluteus muscles (maximus, medius and minimus), fascia lata, piriformis, quadratus femoris, obturator internus and obturator externus.

Furthermore, distinct classification was made for the iliacus, iliopsoas, psoas, abdominal oblique, rectus abdominis, multifidus, femur, iliac and background within the image. CT scan starting slices varied (L3 or L4) among subjects, preventing complete imaging of the iliacus, iliopsoas, psoas, abdominal oblique, rectus abdominis, and multifidus muscles. Due to these inconsistencies, these muscles were excluded from detailed analysis and grouped under ‘Else’. The study primarily examined muscle disparities in the thigh region. An illustrative example of these classifications has been presented in Fig. 2.

Image pre-processing

In the initial phase of pre-processing, we implemented various heuristic techniques to augment the performance of deep learning in visual tasks. The first step involved scaling the intensity range of the CT scan images from −57 to 164, a measure aimed at amplifying the differentiation of individual muscle tissues within the scans (Engelke et al., 2018; Masoudi et al., 2021). Subsequent to this adjustment, we applied a contrast modification using a gamma value of two to further enhance the image clarity. The modified image, characterized by its intensified contrast and reduced metal artifacts, is presented in Fig. 3. To focus on the relevant information, the image was then meticulously cropped to encompass only foreground region.

Deep learning method of automatic muscle segmentation

Our study focuses on the essential task of semantic segmentation, a key component in computer vision for quantifying tasks, particularly crucial in medical imaging. This process aims to delineate and categorize distinct regions of interest inside an imaging (Wang et al., 2022). In terms of muscle segmentation, semantic segmentation necessitates assigning each voxel in the image to a specific category, such as muscle tissue, bone tissue or background. The complexities of this task lie in the need to accurately capture intricate details and variations in the images. This includes accommodating differences in patient’s size, position and tissue textures, all while navigating noise and other imaging artifacts. The ambiguity of muscle tissue in CT scans compounds this challenge, demanding precise voxel-level segmentation (Hiasa et al., 2019). To address this, we employed the architecture of the UNETR model, as depicted in Fig. 4 (Hatamizadeh et al., 2021).

The UNETR model capitalizes on the capabilities of transformers, which have proven to be extraordinarily effective in the timeseries domain, including natural language processing (NLP). By reconceptualizing the task of 3D medical image segmentation as a sequence, the UNETR model utilizes a transformer as the encoder to assimilate sequence representations of the input volume, capturing global multi-scale information. The encoder adopts a U-shaped design, reflective of the original U-net architecture, renowned for its efficacy in biomedical image segmentation (Ronneberger, Fischer & Brox, 2015). This architectural design empowers the model to grasp both high-level contextual insights and granular spatial details, rendering it particularly suitable for individual thigh muscle segmentation. The trained segmentation model shows the segmentation result in Fig. 5.

Results

We selected patients from the hip fracture cohort at Gyeongsang National University Hospital, consisting of six individuals with femoral neck fractures and 12 with intertrochanteric fractures. These patients were further grouped by gender: in the femoral neck fracture category, there were two females and four males, while in the intertrochanteric fracture category, there were 10 females and two males. Table 1 illustrates the average percentages of individual thigh muscle volumes prior to surgical operations. Notably, the vastus muscle group (including lateralis, intermedius, and medialis), along with the adductor magnus and gluteus maximus muscles, constitute significant portions of the thigh muscle composition.

In Figs. 6 and 7, it is evident that there are notable differences in the loss of vastus intermedius and gluteus maximus between the group with femoral neck fractures and the group with intertrochanteric fractures, across both genders. The patterns indicate that the greater the volume contribution of a larger muscle to the thigh, the higher the likelihood of observing differences in muscle atrophy.

Also as detailed in Table 2, prior to making any adjustments, the average total muscle volume loss for females with femoral neck fractures was −83 cm³, compared to −97.4 cm³ for intertrochanteric fractures. For males, the average total muscle volume loss for femoral neck fractures was −147.2 cm³, while intertrochanteric fractures showed a loss of −178.2 cm³. These results suggest that patients with intertrochanteric fractures may be more susceptible to muscle loss and deformities compared to those with femoral neck fractures.

Table 3 illustrates a comparison between femoral neck fractures and intertrochanteric fractures, focusing on the loss of individual thigh muscle volume. We calculated the difference of individual thigh muscle volume between pre-operative and post-operative. To adjust each participant’s characteristic, we also adjusted the muscle volume, achieved by dividing the volume by the square of the height² (m²). Observations indicate that in the female group, individual muscle volumes showed greater discrepancies in cases of intertrochanteric fractures compared to femoral neck fractures (average loss ratio of FNF: 17.8%, average loss ratio of ITF: 42%). However, in the male group, some major muscles like the sartorius (45.2%), vastus intermedius (39.8%), vastus medialis (30.4%) and gluteus maximus (38.0%) demonstrated particularly pronounced disparities for intertrochanteric fractures when compared to femoral neck fractures.

Additionally, as depicted in Fig. 8, which presents data for subgroups merged by gender, the average disparity ratio of total muscle volume in thigh was found to be 29.8% in the femoral neck fracture group and 40.3% in the intertrochanteric fracture group. Most larger thigh muscles exhibited a greater volume loss in intertrochanteric fractures compared to femoral neck fractures. The average percentage of muscles volume loss in vastus lateralis (FNF: 35.1% vs. ITF: 41.8%), vastus intermedius (FNF: 31.8% vs. ITF: 44.4%), vastus medialis (FNF: 26.1% vs. ITF 34.7%), adductor magnus (FNF: 25.6% vs. ITF: 39.4%), gluteus maximus (FNF: 28% vs. ITF: 37.8%), gluteus medius (FNF: 23.5% vs. ITF: 36.9%), sartorius (FNF: 34.2% vs. ITF: 44.9%), rectus femoris (FNF: 19.7% vs. ITF: 34.5%) and adductor longus (FNF: 23% vs. ITF: 43.8%).

Discussion

Hip fractures, particularly those requiring surgical intervention, often bring about significant traumatic pain (Elboim-Gabyzon, Andrawus Najjar & Shtarker, 2019). This traumatic pain, in turn, could influence muscle strength and exacerbate muscle loss might cause gait abnormalities (Xu et al., 2019; Peres-Ueno et al., 2023). Furthermore, these fractures are predominantly observed in the elderly, linking them closely to issue of sarcopenia (Eguchi et al., 2019; Inoue et al., 2020; Chiang, Kuo & Chen, 2021; Park et al., 2022). By incorporating these various factors, we can form a more comprehensive understanding of the complexities surrounding muscle loss due to hip fractures. This enriched perspective sets the stage for future research and clinical practice (Inan et al., 2005; Oh et al., 2020; Kanaya et al., 2023; Robinson et al., 2023).

Building on this context, our research has unveiled some key findings related to the differential muscle loss experienced in patients with intertrochanteric and femoral neck fractures. The analysis of muscle loss patterns, as detailed in Table 2, reveals a key finding that the average total muscle volume loss is higher in patients with intertrochanteric fractures compared to those with femoral neck fractures. This suggests a greater susceptibility to muscle loss and deformities in intertrochanteric fractures, a concern particularly relevant in the context of sarcopenia after hip fracture and the development of rehabilitation strategies.

Further detailed in Figs. 6 and 7, along with Table 3, are the comparison of individual thigh muscle volume losses between femoral neck and intertrochanteric fractures. This comparison, incorporating a ratio to represent the comparative loss, highlights specific muscles such as the vastus lateralis, adductor longus and gluteus maximus are more adversely affected by volume loss in cases of intertrochanteric fractures in both females and males. Conversely, as depicted in Fig. 8, the sartorius muscle exhibits a contrasting trend, with greater disparities in femoral neck fractures, particularly in the subgroups merged by gender. These variations in individual muscle volume loss are critical for developing customized rehabilitation strategies and interventions following hip fracture surgery, highlighting the importance of fracture-type-specific approaches.

In prior studies on muscle deformation related to types of hip fractures, cross-sectional area (CSA) measurements of the psoas and gluteus medius indicated that the CSA of Gluteus medius was significantly larger in cases of intertrochanteric fractures (ITF) than in femoral neck fractures (FNF) (Yerli et al., 2022). Contrarily, our findings, as shown in Table 3, reveal a marked increase in gluteus medius volume loss in the ITF group but only among females, with an average volume loss of 11.8% for ITF compared to 39% for FNF. Another investigation into the CSA of gluteus medius and minimus found no significant difference (Erinç et al., 2020). However, our research illustrates a notable distinction in the female group, where the average loss of gluteus minimus volume in ITF (34.9%) far exceeded that in FNF (13.4%).

Potential explanation for the greater volume loss observed in ITF compared to FNF could be anatomical; major muscles such as the gluteus and vastus are located in the intertrochanteric region and are likely impacted by neural factors as well. Research has shown that bone mineral density (BMD) is significantly associated with the size of the gluteus maximus (p < 0.001) and the mid-thigh area (Yin et al., 2020). This association might partially elucidate why ITF results in significantly greater volume loss in females compared to FNF. The decrease in bone peak in females, primarily due to changes in estrogen levels, is linked to muscle deformation (Spangenburg et al., 2012; Huo et al., 2015). Concerning hip fractures, the intersection point of peak bone and muscle mass loss marks a period of increased hip fracture fragility (Henry et al., 2004; Ho-Pham et al., 2011; Pasco, Nicholson & Kotowicz, 2012). Upon experiencing a hip fracture, females are often already undergoing osteoporosis, and a lower BMD may accelerate the deformation of hip and thigh muscles.

One major limitation of our research lies in the limited dataset size pertaining to hip fracture patients that we could analyze more precisely for muscle volume on hip fracture types. This limitation prevented us from conducting more advanced statistical methods, such as point-biserial correlation or Wilcoxon signed-rank tests. Moreover, the retrospective design of our current study did not allow for the standardized collection of data regarding follow-up duration. To provide a more comprehensive view, we have included the follow-up duration details in the supplementary section. In our future research endeavors, we aim to prospectively gather data, which will enable us to categorize based on follow-up duration. This approach will facilitate the execution of advanced statistical analyses, thereby enhancing the depth of our findings.

Additionally, we acknowledge the significant importance of not only assessing muscle size but also considering the quality of intermuscular adipose tissue (IMAT). However, our training dataset was subject to certain constraints, primarily due to the labor-intensive nature of the manual annotation process using 3D slicer software. This process entailed the initial identification and delineation of muscle regions within specific classes, followed by the application of a hue threshold to differentiate non-muscle tissue in the imaging. These constraints led us to focus predominantly on muscle volume. Despite this focus, we recognize the need for and value of including IMAT volume in future research. To address this, we are considering two approaches: either modifying our existing dataset or developing a new AI model dedicated to the isolation and detailed analysis of IMAT tissue.

In addition, the presence of high artifacts in post-surgery CT scans, originating from screws and implants, might compromise the accuracy of voxel-level segmentation. These limitations should take into account when interpreting our findings and should be the focus of further research to rectify them. To address the limitations of high artifacts in future work, we plan to employ deep learning based pre-processing method by using Deep Residual U-Net architectures, which show promise in minimizing the impact of metal artifacts on the imaging (Selles et al., 2023).

It is crucial, however, to bear in mind that the specific surgical treatments administered for each type of fracture may have a role in shaping these outcomes. Future investigations should therefore focus on isolating the effects that different surgical methods have on muscle atrophy and functionality. Moving forward, we plan to integrate gait and muscle performance into our future research endeavors. This approach will enable us to assess the impact of surgical interventions on muscle function and recovery, providing valuable insights into their effectiveness in enhancing patient outcomes.

Conclusions

In conclusion, our study specifically focuses on the impact of femoral neck and intertrochanteric hip fractures on muscle volume loss, a key factor in sarcopenia among the elderly. Using our UNETR model for muscle segmentation, we found that intertrochanteric fractures, in particularly, lead to significant muscle volume loss in both genders, as evidenced in our results. This reduction in muscle volume was particularly evident in muscles such as the sartorius, the vastus group and gluteus group. These insights directly inform targeted rehabilitation strategies, aiming to improve recovery and quality of life for hip fracture patients. Our study provides a basis for future research into muscle loss associated with hip fractures, aimed at improving treatment and rehabilitation techniques.

Supplemental Information