Deep-learning-based auto-segmentation of lumbar muscle and vertebral body on proton density fat fraction (PDFF) images for clinical applications of body composition analysis

Introduction

Body composition, referring to the amount and distribution of fat and muscle in the body, is in a dynamic change state influenced by external and internal stimuli (1). Body composition analysis has been increasingly recognized in recent years since a large amount of evidence has demonstrated its clinical relevance in various domains, including nutrition and metabolism, aging, degenerative neuromuscular disease, cancer, and so on. For example, recent research shows that the fat infiltration of muscle tissue is associated with the severity and progression of neuromuscular (2-4) and musculoskeletal disorders (5-8). Moreover, the loss of muscle mass (sarcopenia) and myosteatosis have emerged as biomarkers for predicting the prognosis in numerous cancer patients, particularly after chemotherapy (9-12). Besides, there is also a growing body of evidence linking bone marrow adiposity to osteoporosis and treatment-induced bone marrow damage (13-16).

In the cancer field, myelosuppression, either induced by chemotherapy or radiotherapy, or both, is a major treatment-related life-threatening complication (17-19). Bone marrow is a heterogeneous admixture of red and yellow marrow, with red marrow mainly comprising hematopoietic stem cells that produce erythrocytes, leukocytes, and thrombocytes while yellow marrow largely consists of adipose tissue. Treatment-related myelosuppression can cause an increase in the proportion of yellow marrow, leading to higher fat fractions (FFs) in the bone marrow (20,21). Previously, multiple studies have demonstrated that fat quantification in the bone marrow could be used to monitor or predict the severity of hematopoietic toxicity (22-24). Therefore, accurate quantification of muscle mass, fat infiltration of muscle tissue (myosteatosis), and bone marrow adiposity has become an urgent need in cancer management.

As a non-invasive and non-radioactive imaging modality with superb tissue contrast, magnetic resonance imaging (MRI) serves as an ideal body composition quantification tool that depicts anatomic details and provides quantitative composition information. In particular, with the emergence of chemical shift-encoded water-fat imaging (CSE-WFI), the resultant proton density fat fraction (PDFF) images after correcting confounding factors have served as a reliable and accurate approach to quantify fat content (25-27). Furthermore, CSE-WFI can also output an R2* map, facilitating the identification of iron overload (28,29). Therefore, with just one CSE-WFI sequence acquired in the upper abdomen within a breath-hold (~14 s), the volume of muscle, subcutaneous and visceral adipose tissue, and composition of bone marrow could be simultaneously evaluated by exploiting the anatomic information and fat quantification information on PDFF maps, which makes this sequence extremely promising for body composition analysis.

However, conventionally, quantification of fat content within target organs or lesions relies on manual placement of regions of interest (ROIs). This approach carries inherent inter-observer variability and fails to represent the global fat content across the entire organ or lesion. To acquire a holistic measurement of organ- or lesion-wide fat content, analysts must manually delineate and segment tissue boundaries slice-by-slice, a labor-intensive process that remains prone to inter-operator differences. Numerous prior studies (30-32) have described semi-automated segmentation workflows for vertebral bone marrow and abdominal wall muscle fat quantification using PDFF MRI, yet most existing pipelines exhibit inconsistent performance across scanner vendors and require partial manual correction. These persistent limitations hinder widespread clinical translation, creating demand for a fully automated, vendor-stable, objective segmentation strategy such as the one proposed in the present work.

To facilitate the clinical translation of quantitative PDFF MRI for body composition quantification, the present study only aims to technically develop and cross-vendor validate a fully automated deep learning segmentation pipeline covering L1–3 abdominal wall muscles and lumbar vertebral bodies. We additionally performed preliminary exploratory clinical analyses using this segmentation tool, including L3 vertebral marrow FF comparison between patients with and without chemoradiotherapy-induced myelosuppression, as well as a secondary correlation analysis of automated L1–3 total muscle volume versus BMI. We present this article in accordance with the TRIPOD reporting checklist (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0053/rc).

Methods

Dataset for model building and testing

Forty-seven healthy subjects from Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences (Shenzhen, China) who received the fat analysis and calculation technique (FACT) sequence on a 3.0T scanner (uMR790, Shanghai United Imaging Healthcare, Shanghai, China) were retrospectively obtained and were randomly allocated into a training set (n=41 with 898 PDFF images) and a testing set (n=6 with 94 PDFF images). Additionally, another six subjects (138 PDFF images) and five subjects (83 PDFF images) from Cancer Hospital Chinese Academy of Medical Sciences, Shenzhen Center who received iterative decomposition of water and fat with echo asymmetry and least squares estimation quantification sequence (IDEAL-IQ) on 3.0T scanner (Discovery MR 750W, GE Healthcare, Waukesha, USA) and mDixon Quant sequence on another 3.0T scanner (Ingenia, Philips, Best, Netherlands) were retrospectively obtained for independent testing, respectively (Figure 1).

Figure 1 The process of training and testing the segmentation model using PDFF images and clinical application. BMI, body mass index; FF, fat fraction; IDEAL-IQ, iterative decomposition of water and fat with echo asymmetry and least squares estimation quantification sequence; PDFF, proton density fat fraction; UIH, United Imaging Healthcare.

Imaging technique and analysis

Chemical-shift encoded MR PDFF images was performed on 3.0T scanners from three different vendors, namely uMR 790, MR 750W, and Ingenia, using the sequence names FACT, IDEAL-IQ, and mDixon Quant. The imaging procedures covered the lower abdomen (L1–3), and their detailed sequences and parameters are listed in Table 1. The image quality was validated in-line by a radiologist with more than 10 years of experience. If obvious water-fat swapping was present in PDFF images, another scan was repeated immediately with B₀ re-shimming.

PDFF estimation was performed using a multi-echo gradient recalled echo (GRE)-based water-fat separation method with multi-peak fat spectral modeling and T2* correction. For T2* correction, a signal decay model was incorporated into the water-fat separation process. After the initial field map estimation, the T2* decay was determined using an exhaustive search within a reasonable predefined range, as the estimation of T2* decay is not generally affected by multiple local minima. The estimated T2* value was then used to further refine the field map estimation and improve the accuracy of PDFF reconstruction (33).

Dataset for clinical application

For the clinical application, between January 2021 and July 2022, patients who had previously received chemotherapy and/or radiotherapy and undergone a quantitative IDEAL-IQ sequence on a 3.0T MR were retrospectively reviewed. We recorded the time interval from the completion of the final chemotherapy cycle or radiotherapy to IDEAL-IQ MRI for all enrolled patients. The inclusion criteria entailed: (I) abdominal MRI protocol with IDEAL-IQ sequence performed with PDFF map and R2* map generated; (II) scan range covering the first to the third lumbar vertebra; (III) hematological blood tests conducted within three days of MRI scan and the presence and severity of myelosuppression confirmed and assessed. The exclusion criteria comprised: (I) history of vertebral body fractures or metastasis (4 patients excluded); (II) presence of bone islands in the vertebral body and severe degenerative changes or pathological bone changes such as hematological or metabolic bone disorders (6 patients excluded); (III) insufficient image quality due to artifacts (3 patients excluded).

Laboratory blood tests

Blood routine tests, including white blood cell counts (WBC), absolute neutrophil counts (ANC), hemoglobin (Hgb), and platelets, were collected from each enrolled cancer patient. The patients were classified into two groups: the myelosuppression group (hematologic toxicity levels of 1–4) and the non-myelosuppression group (hematologic toxicity levels of 0) (Table 2). The final grade of myelosuppression was determined based on the most severe level among the four items.

Model building

In this study, the PDFF images were uploaded into a commercial software (Slice Omatic, Magog, Quebec, Canada) to manually label the abdominal wall muscles and vertebral bodies from the L1 to L3 level. The manual labeling was accomplished first by two radiologists independently with more than 5 years of experience in abdominal imaging. To maintain high labeling consistency, all ambiguous contours identified during primary delineation were collectively reviewed and adjudicated by a senior abdominal radiologist with more than 10 years of experience in abdominal imaging, with all disagreements resolved through consensus discussion

In our work, a U-net model with EfficientNet from the Segmentation Module Pytorch library was adopted for the segmentation of the abdominal wall muscle and vertebral body (34). The Adam optimization algorithm was chosen as the optimization function, and the weighted cross-entropy was selected as the loss function in training. The model was trained for 100 epochs with a learning rate of 10-4 and implemented in Pycharm (JetBrains, Prague, Czech Republic) on a desktop computer with CPU Intel Xeon 2.88 GHz, 128.0 GB RAM, NVIDIA GeForce RTX 3080 GPU, and 10.0 GB RAM.

In this study, three metrics were used to evaluate the performance of the segmentation network: Dice similarity coefficient (DSC), precision rate (PR), and recall rate (RR).

DSCi=2∙|Gi∩Ri||Gi|+|Ri|,i=1,2[1]

PRi=|Gi∩Ri||Ri|,i=1,2[2]

RRi=|Gi∩Ri||Gi|,i=1,2[3]

Where i represents abdominal wall muscle (i=1) or the vertebral body (i=2). G represents mask images manually labeled as abdominal wall muscle or the vertebral body and R represents the predicted results output by the network. |G∩R| is the total number of pixels with G(r) = R(r) =1, where r is the coordinate index of G and R, and |G| and |R| are the total number of pixels with G(r)=1 and R(r)=1, respectively. If a solitary organ or structure outside the target (e.g., abdominal aorta, costal cartilage or diaphragmatic foot) was mistaken for a predicted target, we determined it as false-positive segmentation. If the target was not predicted at all, we determined it as false-negative segmentation.

To reduce the risk of overfitting, simulate variations in imaging position, and improve segmentation robustness, data augmentation was applied to the training set. Specifically, the augmentation operations included random horizontal flipping and random rotation within a range of −6° to 6°. Regarding intensity normalization, all input images were PDFF maps, with values ranging from 0% to 100%. Before being used for model training, the PDFF values were linearly transformed to an intensity range of 0–255 to standardize the input scale for the neural network.

With respect to the use of two-dimensional (2D) slice-based training, this strategy was adopted for several reasons. First, 2D training substantially increases the number of training samples by treating each axial slice as an individual training instance, which is particularly beneficial when the number of subjects is limited. Second, the abdominal wall muscles, paraspinal muscles, and lumbar vertebral bodies can be clearly identified on axial PDFF images, allowing accurate manual annotation and model learning at the slice level. Third, compared with three-dimensional (3D) training, the 2D slice-based approach requires fewer computational resources and is less dependent on inter-slice consistency or through-plane resolution.

Statistical analysis

Statistical analyses were conducted using SPSS software (version 25). The normality of continuous variables was tested via the Kolmogorov-Smirnov test. Normally distributed variables were presented using the mean ± standard deviation (SD), while non-normally distributed variables were denoted by the median and interquartile range (IQR). Student-t test was employed to compare normally distributed variables, while the Wilcoxon rank sum test was used for non-normally distributed variables. Patient sex and cancer subtypes were compared using chi-squared or Fisher exact tests. Furthermore, Pearson’s correlation was executed to evaluate the relationship between body mass index (BMI) and L1–3 muscle volume. A two-sided P<0.05 was considered statistically significant. We interpreted the magnitude of Pearson correlation coefficients according to published medical statistical reference (35): absolute r values of 0.00–0.10 indicated negligible correlation, 0.10–0.39 weak correlation, 0.40–0.69 moderate correlation, 0.70–0.89 strong correlation, and 0.90–1.00 very-strong correlation.

Ethical consideration

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences (No. SIAT-IRB-210115-H0551) and individual consent for this retrospective analysis was waived.

Results

Clinical characteristics of the enrolled patients for clinical validation

A total of 51 patients were enrolled for the clinical validation, with 27 patients allocated to the myelosuppressive group (12 males and 15 females; mean age, 56±11 years) and 24 patients to the non-myelosuppressive group (12 males and 12 females; mean age, 52±11 years). The median interval from chemoradiotherapy to MRI scan was 12 days (IQR, 6–19 days). No significant differences were observed between the two groups in terms of sex, age, and BMI (P=0.69, 0.19, and 0.10, respectively). In the myelosuppressive group, most patients were classified as grade Ⅱ and grade Ⅲ, accounting for 63% and 33%, respectively. In addition, liver cancer was the most common type of cancer in all patients (n=18). The clinical characteristics of both groups are summarized in Table 3.

Segmentation model building and evaluation

Muscle and vertebral body segmentation

In the testing set, 315 abdomen PDFF images were obtained from 17 subjects through United Imaging Healthcare (UIH) (n=6 with94 PDFF images), GE (n=6 with 138 PDFF images), and Philips (n=5 with 83 PDFF images) scanners. Regarding the number of slices per examination covering L1–3, the UIH cohort had (16±2) slices, the GE cohort had (23±1) slices (6 subjects), and the Philips cohort had (17±2) slices. Overall, the mean DSCs in all the testing datasets were 0.831±0.056 for the abdominal wall muscle and 0.928±0.029 for the vertebral body. Analysis of each vendor revealed that the highest DSCs were obtained on PDFF images acquired from UIH (0.861±0.051 and 0.932±0.027), followed by Philips (0.831±0.043 and 0.930±0.0.029) and GE (0.811±0.057 and 0.925±0.030) both for segmenting abdominal wall muscle and vertebral body (Figures 2,3 & Table 4). Notably, the sequence of effectiveness in terms of PR followed the same trend as in the DSC. However, the model had the highest RR on PDFF images acquired from Philips, followed by UIH and GE.

Figure 2 Representative segmentation results of our model. (A-D) The identical L3 transverse slice (Dice coefficients =0.931 for abdominal wall muscle and 0.924 for vertebral body); (E-H) the matched L2 transverse slice (Dice coefficients =0.890 for abdominal wall muscle and 0.948 for vertebral body). All segmented masks delineate abdominal wall skeletal muscle rather than subcutaneous fat.

Figure 3 Representative examples of false-positive segmentation results obtained on test images. (A-D) The abdominal aorta (yellow arrows) was incorrectly segmented as abdominal wall muscle; (E-H) Costal cartilages (yellow arrows) on both sides of the sternum above the xiphoid process were misidentified as muscle tissue; (I-L) Diaphragmatic crura (yellow arrows) were erroneously included within the abdominal wall muscle.

The testing set included 315 slices, of which 37 slices exhibited false-positive segmentation, including misidentification of the abdominal aorta as the abdominal wall muscle (n=13), misidentification of the costal cartilage on both sides of the sternum at the level above the xiphoid process as muscle (n=13), and the inclusion of the diaphragmatic foot as abdominal wall muscle (n=11) (Figure 3). No false-negative segmentations were identified for muscle segmentation in the testing set. As for vertebral body segmentation, no PDFF images with false-positive and false-negative segmentations were identified in the testing set.

To further quantitatively validate the volumetric accuracy of our segmentation model, we performed a paired t-test to compare automatically derived muscle volumes against manual reference volumes across all 315 axial slices acquired from 17 subjects scanned on three different MRI platforms. The mean slice-wise muscle volume measured by the deep learning model was 92.9 mL, while the corresponding manual segmentation yielded a mean volume of 91.6 mL. The paired t-test revealed no statistically significant difference between the two measurements (P=0.07, 95% CI: −0.10 to 2.79).

Clinical application of the segmentation model

IDEAL-IQ imaging data were acquired from all patients, and the mean FF of the L3 vertebral body was measured for each patient based on automatic segmentation. The mean FF value of the myelosuppression group was significantly higher than that of the non-myelosuppression group (57.17±9.12 vs. 51.18±11.26, P=0.04, 95% CI: −11.73 to −0.25), as illustrated in Figure 4A. Additionally, a moderate positive correlation was observed between L1–3 muscle volume and BMI for all patients (r=0.44, P=0.01), as shown in Figure 4B.

Figure 4 Statistical plots of quantitative metrics derived from patients’ vertebral body and abdominal wall muscle analysis.

Discussion

This study developed and cross-vendor validated a U-Net-based deep learning pipeline for automated segmentation of L1–3 abdominal wall muscles and lumbar vertebral bodies on PDFF maps, with the primary research objective restricted to evaluating the technical robustness and generalizability of the segmentation model rather than validating novel clinical body composition biomarkers. We performed two limited exploratory secondary analyses using the automated segmentation outputs: we compared L3 vertebral marrow FFs between patients with and without chemoradiotherapy-related myelosuppression, and we preliminarily assessed the linear correlation between automated total L1–3 abdominal muscle volume and BMI. The segmentation model achieved consistent, robust segmentation performance across three MRI vendors for both muscle and vertebral tissues. Our exploratory clinical analyses revealed significantly elevated L3 vertebral marrow FF in patients with myelosuppression, alongside a moderate positive correlation between automated L1–3 total muscle volume and patient BMI. We stress that this L1–3 volumetric correlation analysis is only a technical demonstration of our pipeline’s volumetric quantification capacity and cannot be interpreted as evidence supporting L1–3 muscle volume as a validated sarcopenia biomarker.

Based on the fully convolutional network, U-net was initially developed for biomedical image segmentation, specially designed to work with fewer training images and to yield more precise segmentations (36). Since then, U-net and its modified variants have been widely used in the segmentation of various medical images with robust performance (37,38). Previously, a study by Zhou et al. developed a U-net-based deep-learning model to segment vertebral bodies on Dixon Images and achieved robust performance (DSC =0.849±0.091) (39). Our U-net-based model consistently showed robust performance in segmenting abdominal wall muscle and lumbar vertebral body on PDFF images with a relatively small training dataset. Unlike their study which used all four types of mid-sagittal Dixon images (water, fat, R2* map, and PDFF map) as the input and vertebral body segmentation as output, we used only axial PDFF maps as input and both muscle and vertebral body segmentation as output. We obtained a similar performance in the vertebral body segmentation, suggesting the PDFF map provides sufficient quantitative anatomical information to support reliable vertebral and abdominal muscle segmentation in our single-input workflow, as its derivation from both water and fat signal components enables it to capture key compositional information without requiring harmonization of multiple image contrasts. In addition, our model showed robust performance with good generalizability in segmenting PDFF maps both from the same vendor and different vendors, which could be attributed to the advantage of using the quantitative PDFF images that are relatively insensitive to the imaging acquisition parameters, hardware configuration, reconstruction algorithm, and field strengths (40,41). Nevertheless, multi-contrast Dixon inputs, including water, fat, and R2* maps, provide complementary anatomical and technical information that may improve muscle boundary delineation and facilitate detection of artifacts, image quality issues, and water-fat swapping. Therefore, although PDFF-only input offers advantages in simplicity and potential cross-vendor robustness, it provides less information for quality control than multi-contrast approaches. Future studies would evaluate hybrid models that combine PDFF with other Dixon contrasts to further improve robustness.

For body composition segmentation, we found the segmentation performance of the vertebral body was better than that of the abdominal wall muscle across three testing datasets from different vendors. One possible explanation is that segmenting abdominal wall muscle is more challenging due to its complex composition, polymorphism at different levels, and adjacent interfering structures with similar FF values on the PDFF images. For instance, the diaphragmatic foot adjacent to the vertebra was often misregarded as muscle in our study. In addition, due to similar FF values, some structures, such as the abdominal aorta and costal cartilage, were occasionally mistakenly taken as the abdominal wall muscles, leading to poorer segmentation performance. In contrast, the vertebral body can be more easily segmented with high accuracy, potentially attributable to its simple invariable shape, fixed position in the image, and high contrast between the vertebral body and adjacent structures. Therefore, in future studies, the model should be further improved to reduce the false positive rate in segmenting muscle by exploiting additional quantitative images such as R2* images.

The present study also demonstrated a statistically significant difference in the FF between the myelosuppressive group and the non-myelosuppressive group, which was consistent with those of the previous studies (13,42,43). For example, Carmona et al. (13) used PDFF maps to assess the changes in vertebral bone marrow FF during chemoradiotherapy and detected that the bone marrow mean radiation dose received was associated with a 0.43% per Gy increase in PDFF(%). Similarly, Bolan et al. (42) used water-fat MRI to assess bone marrow fat content changes in patients with gynecologic malignancies pre- and post-chemoradiotherapy and observed an increase in FF marrow of vertebral body at the L4 level from baseline to 6 months post-treatment. Our study is a retrospective study with a relatively small sample-size dataset for training, but it preliminarily tested the feasibility of using automatically segmented vertebral bodies on the PDFF maps to assess the bone marrow fat concentration changes. Large multi-cohort deep learning tools (e.g., UK Biobank, NAKO) perform whole-body body composition segmentation via multi-contrast Dixon MRI for population metabolic and aging research (44,45). By contrast, our PDFF-only single-breath-hold model focuses on chemoradiotherapy-treated cancer patients to assess treatment-induced myelosuppression using vertebral marrow FF. While our cohort is relatively small, we validated model generalization across three MRI vendors. Future large prospective multi-center studies on cancer patients, following the large-cohort design of UK Biobank and NAKO, are needed to further verify our pipeline’s clinical value.

In this study, the deep-learning-based auto-segmentation model demonstrated robust performance in segmenting abdominal wall muscle and vertebral body on PDFF maps and may be a promising tool for body composition analysis.

Artificial intelligence tools were used for assistance in identifying areas for manuscript revision based on reviewer comments.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0053/rc

Data Sharing Statement: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0053/dss

Peer Review File: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0053/prf

Funding: This study has received funding by Shenzhen High-level Hospital Construction Fund; Shenzhen Clinical Research Center for Cancer (No. [2021] 287); National Key Research and Development Program of China, Grant/Award Number (No.2022YFA1004203); Key Laboratory Project of Guangdong Province, Grant/Award Number (No. 2020B1212060051); Scientific Instrument Innovation Team of the Chinese Academy of sciences, Grant/Award Number (GJJSTD20180002); Shenzhen Municipal Scientific Program, Grant/Award Number (JCY20200109110612375); National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital & Shenzhen Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Shenzhen (SZ2020ZD005).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2026-0053/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences (No. SIAT-IRB-210115-H0551) and individual consent for this retrospective analysis was waived.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Weston AD, Korfiatis P, Kline TL, et al. Automated Abdominal Segmentation of CT Scans for Body Composition Analysis Using Deep Learning. Radiology 2019;290:669-79. [Crossref] [PubMed]
2. Gadermayr M, Disch C, Müller M, et al. A comprehensive study on automated muscle segmentation for assessing fat infiltration in neuromuscular diseases. Magn Reson Imaging 2018;48:20-6. [Crossref] [PubMed]
3. Dahlqvist JR, Widholm P, Leinhard OD, et al. MRI in Neuromuscular Diseases: An Emerging Diagnostic Tool and Biomarker for Prognosis and Efficacy. Ann Neurol 2020;88:669-81. [Crossref] [PubMed]
4. Reyngoudt H, Marty B, Boisserie JM, et al. Global versus individual muscle segmentation to assess quantitative MRI-based fat fraction changes in neuromuscular diseases. Eur Radiol 2021;31:4264-76. [Crossref] [PubMed]
5. Kumar D, Karampinos DC, MacLeod TD, et al. Quadriceps intramuscular fat fraction rather than muscle size is associated with knee osteoarthritis. Osteoarthritis Cartilage 2014;22:226-34. [Crossref] [PubMed]
6. Baum T, Yap SP, Dieckmeyer M, et al. Assessment of whole spine vertebral bone marrow fat using chemical shift-encoding based water-fat MRI. J Magn Reson Imaging 2015;42:1018-23. [Crossref] [PubMed]
7. Davison MJ, Maly MR, Adachi JD, et al. Relationships between fatty infiltration in the thigh and calf in women with knee osteoarthritis. Aging Clin Exp Res 2017;29:291-9. [Crossref] [PubMed]
8. Kalichman L, Carmeli E, Been E. The Association between Imaging Parameters of the Paraspinal Muscles, Spinal Degeneration, and Low Back Pain. Biomed Res Int 2017;2017:2562957. [Crossref] [PubMed]
9. Blauwhoff-Buskermolen S, Versteeg KS, de van der Schueren MA, et al. Loss of Muscle Mass During Chemotherapy Is Predictive for Poor Survival of Patients With Metastatic Colorectal Cancer. J Clin Oncol 2016;34:1339-44. [Crossref] [PubMed]
10. Hayashi N, Ando Y, Gyawali B, et al. Low skeletal muscle density is associated with poor survival in patients who receive chemotherapy for metastatic gastric cancer. Oncol Rep 2016;35:1727-31. [Crossref] [PubMed]
11. Lee CM, Kang J. Prognostic impact of myosteatosis in patients with colorectal cancer: a systematic review and meta-analysis. J Cachexia Sarcopenia Muscle 2020;11:1270-82. [Crossref] [PubMed]
12. Sheean P, Gomez-Perez S, Joyce C, et al. Myosteatosis at diagnosis is adversely associated with 2-year survival in women with estrogen receptor-negative metastatic breast cancer. Breast Cancer Res Treat 2021;190:121-32. [Crossref] [PubMed]
13. Carmona R, Pritz J, Bydder M, et al. Fat composition changes in bone marrow during chemotherapy and radiation therapy. Int J Radiat Oncol Biol Phys 2014;90:155-63. [Crossref] [PubMed]
14. Hirschfeld HP, Kinsella R, Duque G. Osteosarcopenia: where bone, muscle, and fat collide. Osteoporos Int 2017;28:2781-90. [Crossref] [PubMed]
15. Dieckmeyer M, Ruschke S, Rohrmeier A, et al. Vertebral bone marrow fat fraction changes in postmenopausal women with breast cancer receiving combined aromatase inhibitor and bisphosphonate therapy. BMC Musculoskelet Disord 2019;20:515. [Crossref] [PubMed]
16. Sollmann N, Löffler MT, Kronthaler S, et al. MRI-Based Quantitative Osteoporosis Imaging at the Spine and Femur. J Magn Reson Imaging 2021;54:12-35. [Crossref] [PubMed]
17. Sacks EL, Goris ML, Glatstein E, et al. Bone marrow regeneration following large field radiation: influence of volume, age, dose, and time. Cancer 1978;42:1057-65. [Crossref] [PubMed]
18. Li J, Law HK, Lau YL, et al. Differential damage and recovery of human mesenchymal stem cells after exposure to chemotherapeutic agents. Br J Haematol 2004;127:326-34. [Crossref] [PubMed]
19. Georgiou KR, Foster BK, Xian CJ. Damage and recovery of the bone marrow microenvironment induced by cancer chemotherapy - potential regulatory role of chemokine CXCL12/receptor CXCR4 signalling. Curr Mol Med 2010;10:440-53. [Crossref] [PubMed]
20. Fliedner TM, Graessle D, Paulsen C, et al. Structure and function of bone marrow hemopoiesis: mechanisms of response to ionizing radiation exposure. Cancer Biother Radiopharm 2002;17:405-26. [Crossref] [PubMed]
21. Platoff R, Villalobos MA, Hagaman AR, et al. Effects of radiation and chemotherapy on adipose stem cells: Implications for use in fat grafting in cancer patients. World J Stem Cells 2021;13:1084-93. [Crossref] [PubMed]
22. Orlandini GE, Ruggiero L, Gulisano M, et al. Magnetic resonance (MR) evaluation of bone marrow in vertebral bodies. Arch Ital Anat Embriol 1991;96:93-100. [PubMed]
23. Lee EYP, Perucho JAU, Vardhanabhuti V, et al. Intravoxel incoherent motion MRI assessment of chemoradiation-induced pelvic bone marrow changes in cervical cancer and correlation with hematological toxicity. J Magn Reson Imaging 2017;46:1491-8. [Crossref] [PubMed]
24. Werner S, Krauss B, Horger M. Dual-energy CT based monitoring of treatment-induced bone marrow changes in lung cancer patients: preliminary results. Quant Imaging Med Surg 2022;12:1871-81. [Crossref] [PubMed]
25. Idilman IS, Aniktar H, Idilman R, et al. Hepatic steatosis: quantification by proton density fat fraction with MR imaging versus liver biopsy. Radiology 2013;267:767-75. [Crossref] [PubMed]
26. Kim KY, Song JS, Kannengiesser S, et al. Hepatic fat quantification using the proton density fat fraction (PDFF): utility of free-drawn-PDFF with a large coverage area. Radiol Med 2015;120:1083-93. [Crossref] [PubMed]
27. Hernando D, Sharma SD, Aliyari Ghasabeh M, et al. Multisite, multivendor validation of the accuracy and reproducibility of proton-density fat-fraction quantification at 1.5T and 3T using a fat-water phantom. Magn Reson Med 2017;77:1516-24. [Crossref] [PubMed]
28. Corrias G, Erta M, Sini M, et al. Comparison of Multimaterial Decomposition Fat Fraction with DECT and Proton Density Fat Fraction with IDEAL IQ MRI for Quantification of Liver Steatosis in a Population Exposed to Chemotherapy. Dose Response 2021;19:1559325820984938. [Crossref] [PubMed]
29. Zerunian M, Pucciarelli F, Masci B, et al. Updates on Quantitative MRI of Diffuse Liver Disease: A Narrative Review. Biomed Res Int 2022;2022:1147111. [Crossref] [PubMed]
30. Sollmann N, Dieckmeyer M, Schlaeger S, et al. Associations Between Lumbar Vertebral Bone Marrow and Paraspinal Muscle Fat Compositions-An Investigation by Chemical Shift Encoding-Based Water-Fat MRI. Front Endocrinol (Lausanne) 2018;9:563. [Crossref] [PubMed]
31. Burian E, Rohrmeier A, Schlaeger S, et al. Lumbar muscle and vertebral bodies segmentation of chemical shift encoding-based water-fat MRI: the reference database MyoSegmenTUM spine. BMC Musculoskelet Disord 2019;20:152. [Crossref] [PubMed]
32. Somasundaram A, Wu M, Reik A, et al. Evaluating Sex-specific Differences in Abdominal Fat Volume and Proton Density Fat Fraction at MRI Using Automated nnU-Net-based Segmentation. Radiol Artif Intell 2024;6:e230471. [Crossref] [PubMed]
33. Yu H, Shimakawa A, McKenzie CA, et al. Multiecho water-fat separation and simultaneous R2* estimation with multifrequency fat spectrum modeling. Magn Reson Med 2008;60:1122-34. [Crossref] [PubMed]
34. R P, M JPP, J S N. Brain tumor segmentation using multi-scale attention U-Net with EfficientNetB4 encoder for enhanced MRI analysis. Sci Rep 2025;15:9914. [Crossref] [PubMed]
35. Wang Q, Ye T, Chen HL, et al. Correlation between intensity modulated radiotherapy and bone marrow suppression in breast cancer. Eur Rev Med Pharmacol Sci 2016;20:75-81. [PubMed]
36. Zunair H, Ben Hamza A. Sharp U-Net: Depthwise convolutional network for biomedical image segmentation. Comput Biol Med 2021;136:104699. [Crossref] [PubMed]
37. Kihira S, Mei X, Mahmoudi K, et al. U-Net Based Segmentation and Characterization of Gliomas. Cancers (Basel) 2022;14:4457. [Crossref] [PubMed]
38. Anand V, Gupta S, Koundal D, et al. Modified U-NET Architecture for Segmentation of Skin Lesion. Sensors (Basel) 2022;22:867. [Crossref] [PubMed]
39. Zhou J, Damasceno PF, Chachad R, et al. Automatic Vertebral Body Segmentation Based on Deep Learning of Dixon Images for Bone Marrow Fat Fraction Quantification. Front Endocrinol (Lausanne) 2020;11:612. [Crossref] [PubMed]
40. Hu HH, Börnert P, Hernando D, et al. ISMRM workshop on fat-water separation: insights, applications and progress in MRI. Magn Reson Med 2012;68:378-88. [Crossref] [PubMed]
41. Yokoo T, Browning JD. Fat and iron quantification in the liver: past, present, and future. Top Magn Reson Imaging 2014;23:73-94. [Crossref] [PubMed]
42. Bolan PJ, Arentsen L, Sueblinvong T, et al. Water-fat MRI for assessing changes in bone marrow composition due to radiation and chemotherapy in gynecologic cancer patients. J Magn Reson Imaging 2013;38:1578-84. [Crossref] [PubMed]
43. Wang C, Qin X, Gong G, et al. Correlation between changes of pelvic bone marrow fat content and hematological toxicity in concurrent chemoradiotherapy for cervical cancer. Radiat Oncol 2022;17:70. [Crossref] [PubMed]
44. Jung M, Reisert M, Rieder H, et al. Body Composition in the General Population: Whole-body MRI-derived Reference Curves from Over 66 000 Individuals. Radiology 2026;319:e251939. [Crossref] [PubMed]
45. Graf R, Platzek P, Riedel EO, et al. VIBESegmentator: full body MRI segmentation for the NAKO and UK Biobank. Eur Radiol 2026;36:2548-62. [Crossref] [PubMed]