Introduction
Patients with multiple myeloma (MM) are frequently exposed to prolonged, multiline and multidrug treatments, factors that may substantially influence overall health, physical reserve, and physiological resilience. Despite this, commonly used assessments of treatment effects in this population rely largely on indirect measures such as laboratory biomarkers of organ function (eg, creatinine), patient-reported symptoms such as fatigue or pain, or global functional status. Many of these surrogate or global markers lack sensitivity and/or specificity regarding the early stages of underlying organ changes, due to methodological deficits, but also due to frequent physiological compensation. This constitutes a significant barrier to early diagnosis and prevention of organ treatment side effects, for example through selection of less toxic therapies or drug schedules. Contemporary imaging techniques, which are increasingly used for imaging the tumor inside and outside the bone marrow in patients with myeloma, provide an opportunity to move beyond these surrogate markers. Imaging of organs that are inevitably captured by whole-body scans can be opportunistically interrogated to derive direct, quantitative measures of body composition. In an era of rapidly expanding clinical trials, approaches that enable systematic assessment and comparison of treatment-related changes in body composition are likely to become increasingly important.
Data on sensitivity, quality of life, cost effectiveness, and sparing of radiation exposure have shaped national and international guidelines to include noninvasive whole-body magnetic resonance imaging (WBMRI) for MM diagnosis. 1 , 2 , 3 More recently evidence on the application of WBMRI to assess residual disease has also seen WBMRI emerge as a modality for response assessment. 4 , 5
The standardization of WBMRI acquisition (Myeloma Response Assessment and Diagnosis System) in line with other oncological WBMRI protocols 6 , 7 , 8 together with recent improvements in artificial intelligence (AI) methodologies for automated organ and tissue segmentation presents an opportunity to complement accurate disease assessments with body composition metrics. The need to identify patients who are at risk of not achieving maximum benefit from treatment has been highlighted as a priority, 9 however the contribution of measures of body composition toward this goal is poorly defined.
Sarcopenia is formally recognized as a muscle disease defined in a European consensus as loss of muscle strength, muscle quantity/quality, with or without loss of physical performance. 10 Measured on a single computed tomography (CT) performed within 6 months of diagnosis at the level of the L3 vertebral body, sarcopenia has been shown to adversely affect survival in patients with newly diagnosed myeloma, 11 with a much more established body of evidence for a similar effect in other cancer types. 12 Elsewhere, both high and low extremes of body mass index (BMI) have been reported to be associated with inferior outcomes. 13
Although sarcopenia may be assessed from quantifying muscle mass in several imaging modalities such as dual-energy X-ray absorptiometry and CT, these modalities are not optimal or used routinely for both myeloma diagnosis and follow-up. Routine longitudinal WBMRI could provide unparalleled opportunistic insights into nondiseased tissue that extends beyond conventional measures of sarcopenia.
The aim of this study is to develop an AI-based pipeline to enable automated quantitative image-derived phenotypes (IDPs) in nondiseased tissue from WBMRI in patients with MM and to perform baseline and longitudinal measurements and exploratory associations with outcomes.
Materials and methods
Trial design
Testing and retraining an AI-based pipeline for automating IDPs
Institutional review board and national ethics committee approval were obtained for testing and retraining of the WBMRI segmentation model (IRAS 436000). This study was conducted in accordance with the Declaration of Helsinki.
A precision phenotyping pipeline (Orcino Health Ltd, London, United Kingdom) for T1-weighted Dixon WBMRI data developed using healthy participants’ data in the UK Biobank (application number 105847) was used. The pipeline mirrors processing that has been described and used in several publications covering its utility when applied to large population studies in a mostly healthy population. 14 , 15 , 16 However, because the pipeline has not been exposed to WBMRI data from patients with MM, testing and retraining were planned.
For the fine-tuning the deep-learning segmentation models, 40 retrospective, routinely acquired WBMRI scans from patients with active myeloma were identified from the Radiology Information System at this institution. The scans were performed any time from 01 January 2018 to 10 July 2024. Scans were selected such that at least half displayed features of background hypercellular marrow and half displayed focal marrow lesions (both features could be present), and 10 with artifacts from metal prostheses because these are imaging findings to which the IDP algorithm had not been exposed previously. These scans were only used to train and evaluate the model; the analysis was performed on a separate cohort (image-guided theranostics in MM [iTIMM], see “Measureing baseline and longitudinal changed in IDP and exploring the relationship with clinical parameters”).
Standardized clinical 6 , 7 , 8 , 17 and UK Biobank 18 WBMRI protocols rely on breath-holds for T1 Dixon scanning to ensure reduction of breathing motion. This standardization therefore also supports translation of this work to data acquired at other sides and in other diseases.
The deep-learning model was applied to 20 data sets to predict segmentations for the body structures outlined in Table 1 . A success criterion for this task was defined on a per-structure basis such that the relevant model inference passed if the mean performance (as measured by the Dice score when the segmentation was compared with a ground-truth, expert-annotated mask) across the 20 data sets exceeded the minimum threshold score shown in Table 1 . These minimum Dice thresholds were based on those established using a broad spectrum of participants the UK Biobank for similar deep-learning models in the literature. 14 A definition of the Dice score is provided in the supplemental Methods . Table 1. Minimum Dice score thresholds to assess deep-learning model performance for organ, tissue, and bone segmentations Structure Functional group Minimum Dice threshold Body cavity Body composition 0.9 Abdominal cavity Skeletal muscle Liver Internal organs 0.85 Spleen Kidneys (left and right) Lungs Heart with major vessels Iliopsoas muscles (left and right)
Table 1.
Minimum Dice score thresholds to assess deep-learning model performance for organ, tissue, and bone segmentations
| Structure | Functional group | Minimum Dice threshold |
| Body cavity | Body composition | 0.9 |
| Abdominal cavity | ||
| Skeletal muscle | ||
| Liver | Internal organs | 0.85 |
| Spleen | ||
| Kidneys (left and right) | ||
| Lungs | ||
| Heart with major vessels | ||
| Iliopsoas muscles (left and right) |
The Likert scale is commonly used to assess quality in medical imaging. 19 Once satisfactory Dice scores were achieved, a Likert quality scale from 1 to 5 ( Table 2 ), was independently performed by 2 imaging specialists (A.D. and R.E.) to rate the performance of organ, tissue, and bone segmentations from the final deep-learning models. The final scores were averaged across structures and assessor. Table 2. Five-point Likert quality scale Score Criteria 5 Excellent Negligible changes, could be used clinically without edits 4 Good Minor changes on <4 slices or changes that would take <1 min to fix 3 Fair Edits needed on multiple slices, but <50% 2 Poor Edits needed on >50% of slices 1 Very poor Structure is poorly predicted, cannot be used
Table 2.
Five-point Likert quality scale
| Score | Criteria | |
| 5 | Excellent | Negligible changes, could be used clinically without edits |
| 4 | Good | Minor changes on <4 slices or changes that would take <1 min to fix |
| 3 | Fair | Edits needed on multiple slices, but <50% |
| 2 | Poor | Edits needed on >50% of slices |
| 1 | Very poor | Structure is poorly predicted, cannot be used |
Each WBMRI scan was segmented, using deep-learning models, into 3-dimensional (3D) structures ( Table 1 ), which were used to generate 11 IDPs: volumes of abdominal subcutaneous adipose tissue (ASAT), visceral adipose tissue (VAT), abdominal skeletal muscle (ASM), liver, spleen, left and right kidneys, left and right iliopsoas muscles, and the heart. Additionally, the relative fat fraction (rFF) in the liver was estimated from the Dixon sequences using the fat-to-water ratio for each voxel in the segmented liver and applying the median. An example MM WBMRI image and resulting segmentations are shown in Figure 1 Figure 1. Schematic of pipeline input (WBMRI) and output (3D segmentations). The pipeline takes WBMRI data as input and returns a set of IDPs after applying several neural network and postprocessing steps.
Measuring baseline and longitudinal changes in IDP and exploring the relationship with clinical parameters
We used data from a single-center, prospective observational study (iTIMM, ClinicalTrials.gov identifier: NCT02403102 ) of WBMRI in patients with newly diagnosis or first-relapsed symptomatic MM who were planned to undergo induction therapy followed by high-dose melphalan and autologous stem cell transplant (ASCT). WBMRI was acquired at baseline (MRI1), after induction therapy (MRI2), and after autograft (MRI3). Patients were not included in the testing and retraining of the pipeline and were in follow-up for progression-free survival (PFS) up to 5 years. Detailed protocols, patient and treatment characteristics, and the superior sensitivity of WBMRI compared with fluorodeoxyglucose positron emission tomography/CT, and the correlation between WBMRI response at 3 months after ASCT and patient outcomes have been previously reported. 4 , 20
In brief, iTIMM included patients aged >18 years with a new diagnosis or first relapse of symptomatic MM who were planned to undergo high-dose melphalan and ASCT. Patients with a diagnosis of another malignancy in the previous 5 years or contraindications to MRI or fluorodeoxyglucose positron emission tomography/CT were excluded. Diagnosis and relapse were defined according to International Myeloma Working Group criteria using laboratory parameters, including bone marrow trephine and assessment of end-organ damage. 21 , 22 , 23
WBMRI was performed in line with the Myeloma Response Assessment and Diagnosis System consensus protocol (T1- and T2-weighted sagittal spine, axial diffusion-weighted, and axial 3D gradient-echo Dixon imaging from head to knees 5 ) and used a MAGNETOM Avanto 1.5T scanner (Siemens Healthineers, Forchheim, Germany).
All imaging and clinical variables (including age and sex) available for measuring associations or building regression models are provided in the supplemental Methods . Patients from the iTIMM study were evaluable for the primary end point of this study if they had a day 100 post-ASCT WBMRI scan and had completed at least 2 years of follow-up for progression. Median follow-up was 42 months. 4 One patient had a missing scan at MRI2.
Statistical analysis
Segmentation model performance
Segmentation model performance was assessed using the Dice score and a Likert scale ( Table 2 ) graded by a consultant radiologist and a senior research radiographer.
IDP association with patient outcomes
All summary statistics, hypothesis tests, regression models, and figures were performed using the R software environment for statistical computing and graphics (version 4.3.3). Associations between measurements at the same scan were assessed by Pearson correlation. Differences between measurements at 2 scans were assessed by paired t tests. The Cox proportional hazards (PH) model was used to estimate univariate and multivariate hazard ratios and 95% confidence intervals for PFS. Associations between PFS and a combination of IDPs and clinical metrics were performed for the following 6 scenarios: 1. Only clinical metrics and IDPs from MRI1 are allowed in the model. 2. Clinical metrics from MRI1 and MRI2, and only IDPs from MRI2 are allowed in the model. 3. Clinical metrics from MRI1, MRI2, and MRI3; and only IDPs from MRI3 are allowed in the model. 4. Clinical metrics from MRI1 and MRI2, IDPs from MRI, and the difference in IDPs between MRI 1 and 2 are allowed in the model. 5. Clinical metrics from MRI1, MRI2, and MRI3; IDPs from MRI1; and the difference between IDPs from MRI 1 and 3 are allowed in the model. 6. Clinical metrics from MRI1, MRI2, and MRI3; IDPs from MRI1; the difference between IDPs from MRI1 and MRI2; and the difference between IDPs from MRI2 and MRI3 are allowed in the model.
All possible combinations of the clinical variables and IDPs were constructed from univariate models in which only 1 variable is included to the multivariate model that includes all variables, and a Cox PH model fit for each combination. For example, >130 000 combinations of variables (and by association Cox PH models) are available for 17 variables. The Bayesian information criterion (BIC) and concordance index (c-index) were used to select the “best” model, because the BIC encourages parsimony in a statistical model 24 and the c-index measures the performance of a time-to-event model. 25 Model selection was performed by balancing low BIC values and high c-index values. Specifically, 3 models were selected from all possible combinations, in which the first model selected had the lowest BIC value with a c-index of >0.60, the second model had a higher c-index value relative to the first model while increasing the BIC value as little as possible, and the third model had a higher c-index value relative to the second model while further increasing the BIC value as little as possible. A detailed example is provided in the supplemental Results . Note that Cox PH models were not selected in a stepwise manner so when variables are in >1 model, with similar model performance, this provides evidence that the variable has a more robust association with PFS.
P values <.05 were considered statistically significant. Log transformation was applied to liver rFF because of its high level of skewness when included in the Cox PH models.
Results
Patient population
In total, 39 WBMRI scans were used for testing and retraining the segmentation models. One scan was discarded because of a technical error in the DICOM (Digital Imaging and Communications in Medicine) data. Twenty scans were used for training, and the remaining 19 scans were used for validation.
Of 105 patients in the iTIMM study, 69 (65.7%) fulfilled inclusion criteria for this study and were included in the analysis (see Figure 2 for CONSORT diagram). Imaging data for 2 patients at MRI3 were corrupted; however, the data at MRI1 and MRI2 were successfully acquired and processed. Therefore, the analyses relating to MRI3 included 67 patients, whereas all other analyses included 69 patients. Figure 2. CONSORT diagram detailing the inclusion criteria for patients recruited to iTIMM who were included in this study. A total of 69 patients for analyses with MRI1 and MRI2 data, and 67 patients with MRI3 data. FDG, fluorodeoxyglucose; LFUP, lost to follow-up; PET, positron emission tomography; RT, radiotherapy; SD, standard deviation; TX, treatment.
Model segmentation quality metrics
After the first round of retraining, the segmentation models exceeded the defined success criteria, for each structure. The mean Dice scores and the Likert scores are shown in Table 3 . The mean Dice score across all structures was 0.916, and the mean Likert score across all structures was 4.58, passing the thresholds set out at the onset. Therefore, further validation scans were not required, and our model was ready to be applied to the iTIMM scans. Table 3. Mean Dice scores and mean Likert scores (2 assessors) for model segmentation performance Structure Mean Dice score Mean Likert score Abdominal cavity 0.960 4.24 Body cavity 0.990 4.87 Total muscle 0.934 4.24 Spleen 0.882 4.51 Liver 0.945 4.48 Right kidney 0.924 4.81 Left kidney 0.944 4.81 Right iliopsoas 0.892 4.49 Left iliopsoas 0.895 4.73 Heart 0.891 4.57
Table 3.
Mean Dice scores and mean Likert scores (2 assessors) for model segmentation performance
| Structure | Mean Dice score | Mean Likert score |
| Abdominal cavity | 0.960 | 4.24 |
| Body cavity | 0.990 | 4.87 |
| Total muscle | 0.934 | 4.24 |
| Spleen | 0.882 | 4.51 |
| Liver | 0.945 | 4.48 |
| Right kidney | 0.924 | 4.81 |
| Left kidney | 0.944 | 4.81 |
| Right iliopsoas | 0.892 | 4.49 |
| Left iliopsoas | 0.895 | 4.73 |
| Heart | 0.891 | 4.57 |
Longitudinal changes in IDPs
On analysis of the iTIMM scans, we found changes in most IDPs when averaging across all patients ( Figure 3 ). ASM volume decreased between baseline (MRI1) and MRI2 ( P < .001) and did not change between MRI2 and MRI3. ASAT and VAT volume increased between MRI1 and MRI2 ( P < .001) and decreased between MRI2 and MRI3 ( P < .001). Given the relative contributions of ASM, ASAT, and VAT to total body mass, it is not surprising that BMI did not change between MRI1 and MRI2 and decreased between MRI2 and MRI3 ( P < .001). Liver rFF increased between MRI1 and MRI2 ( P < .005) as well as overall between MRI1 and MRI3 ( P < .001). There was a small decrease in liver rFF between MRI2 and MRI3, although it did not reach statistical significance ( P = .33). Figure 3. Changes in IDPs and BMI between MRI time points for all patients. The circle represents the mean change, and the line represents the standard error of the mean. SD, standard deviation.
With respect to internal organs, heart volume decreased between MRI1 and MRI2 ( P < .001) and did not change between MRI2 and MRI3. However, because the Dice scores for cardiac segmentation were lowest, possibly due to cardiac motion and variable inclusion of aorta and vena cava, the clinical significance of this finding is uncertain. Left kidney and right kidney volumes decreased between MRI2 and MRI3 ( P = .029 and P = .023, respectively). The left and right iliopsoas muscle volumes decreased between MRI1 and MRI3 ( P < .001 and P < .001, respectively) and did not change between the MRI2 and MRI3, following the pattern of ASM volume. There was no statistically significant change in liver or spleen volumes between any of the time points.
Association of clinical measurements and IDPs with outcome
Substantial correlations were found between ASM and left/right iliopsoas muscles (ρ = 0.81 and ρ = 0.85, respectively) at MRI1 ( supplemental Figure 1 ). The iliopsoas muscles were excluded from the analysis because of their high correlation with ASM and the fact that they are included within the definition of ASM ( supplemental Figure 2 ). This exclusion applies to all time points.
We found body composition and several internal organ volumes to be highly associated with patient outcome in a Cox PH multivariable regression model. PFS was selected for analysis because it contained a greater number of events (n = 45) compared with overall survival (n = 13). The variables for the 3 “best” Cox PH models for PFS are presented in supplemental Table 7 for the 6 different scenarios. All IDPs represent volumes, except for rFF in the liver, which is expressed as a percentage.
Main findings across the 6 modeling scenarios
At baseline (MRI1), a higher ASM and ASAT were associated with longer PFS in all 3 models. Kaplan-Meier curves were constructed to illustrate the univariate association between binarized versions of ASM and ASAT volume with PFS using thresholds of 2.63 L/m 2 and 1.34 L/m 2 , respectively ( Figure 4 ). The best possible threshold, producing the smallest P value in the Kaplan-Meier curve between the 2 groups, was selected to create the dichotomized variables. This analysis was very exploratory and added to provide a visual representation of the relationship between the IDP and PFS. Clear separation in PFS was observed between patients with higher volumes of ASM or ASAT compared with patients with lower volumes. For comparison, no clear separation in PFS was observed for binarized BMI (binarized at 26.7 kg/m 2 ). Figure 4. Kaplan-Meier survival curves by tissue volume. Progression-free survival stratified by high versus low values of (A) abdominal skeletal muscle and (B) abdominal subcutaneous adipose tissue volumes.
Higher BMI at MRI1 was associated with shorter PFS and higher VAT with longer PFS, but only in 1 of 3 models. Note that BMI is highly correlated with body composition IDPs ASM and ASAT in this model; therefore, interpretation should be cautious. Looking at the regression coefficients, it is most likely that BMI in isolation is associated with longer PFS. At MRI3, higher ASM was associated with longer PFS in 2 of 3 models, and higher liver fat fraction was associated with shorter PFS in only 1 of 3 models.
When changes in IDPs between MRI1 and MRI2 were introduced, there was an association between higher ASAT at baseline (MRI1) with longer PFS in 2 of 4 models; higher ASM at baseline (MRI1) was associated with longer PFS in 2 of 3 models and increased liver rFF between MRI1 and MRI2 was associated with longer PFS in 2 of 3 models.
When changes in IDPs between MRI1 and MRI3 were introduced into the model, an increase in liver rFF over time was associated with longer PFS in all 3 models; an increase in VAT over time was associated with shorter PFS in all 3 models; and higher ASM at baseline (MRI1) was associated with longer PFS in 2 of 3 models. An increase in BMI over time and a decrease in ASM over time were associated with longer PFS, but the 2 variables are highly correlated, meaning they most likely reduce the overall effect of body composition in this scenario.
Full details are presented in supplemental Tables 1-6 .
Discussion
Our results extend the clinical utility of WBMRI beyond lesion detection and response assessment by showing that automated, robust segmentation of nondiseased tissues is feasible and delivers IDPs that change with treatment and are associated to outcomes. The high average Dice and Likert scores indicate that models trained on population cohorts (UK Biobank) can be successfully adapted to patients with MM. Imaging used for diagnosis and imaging-based minimal residual disease and response assessment are increasingly central to myeloma care, and adding body composition metrics to WBMRI workflow leverages the same acquisition without additional burden to patients.
Despite therapeutic advances, management of MM remains lengthy and burdensome, with patients commonly experiencing morbidity due to both disease-related sequelae and the adverse effects of therapy. Accordingly, this study demonstrated significant reductions in renal volume at posttransplant follow-up; a reduction in ASM after induction chemotherapy; increased ASAT and VAT after induction chemotherapy and a decrease after transplantation; and an overall increase in liver rFF.
The association of greater ASM with longer PFS aligns with literature linking sarcopenia to poorer outcomes across cancers and with previous studies in myeloma suggesting adverse survival with low muscle mass. 11 , 26 However, not all CT studies of sarcopenia have shown an association with survival. 27 , 28 , 29 , 30 These discordant results may, in part, be explained by variability in the populations and differences in methodologies and thresholds used to define sarcopenia.
Greater ASAT was also commonly associated with longer PFS in survival models, which aligns with previously reported findings in other cancers, which may reflect advantages of better baseline reserves although altered immune responses have also been proposed. 31 In a study of CT-derived measures, patients with disease progression/death during follow-up had significantly lower subcutaneous adipose tissue than those with progression-free follow-up. 32 A study of BMI and survival in patients with newly diagnosed MM suggested that underweight and severe obesity are associated with worse survival 13 although the “obesity paradox” is controversial because BMI does not differentiate lean vs adipose tissue compartments (ie, body composition). 33 The complex associations between ASAT, VAT, and BMI in our models and in a study by Groß et al may be explained because BMI is a crude measure, which can mask divergent effects of lean and fat compartments. 34 Longitudinal loss of muscle volume has also been reported on CT studies 35 , 36 and the serial CT study from Kylies et al also reported that loss of muscle volume was an independent predictor of reduced survival.
Increasing VAT over time was associated with poorer PFS. This is concordant with CT-derived metrics of VAT, which have been shown to be associated with inferior response to a bortezomib-based induction therapy, whereas BMI was not. 34 In that study an interference of VAT with the mode of action of bortezomib was postulated.
Changes in body composition in patients with MM are likely to be profoundly influenced by steroid administration, which conversely to most solid cancers, is a cornerstone of induction therapy. Steroids not only increase appetite but also lead to increases in VAT, ASAT, liver rFF, and decreases in ASM. Associations between baseline liver rFF and increases over time may therefore relate to the effects of steroid therapy.
The findings of this study are an important step toward use of automated IDPs from WBMRI as noninvasive biomarkers for risk stratification, treatment planning, and survivorship care. Early identification of patients at risk through body composition monitoring could provide an opportunity for timely interventions, potentially slowing disease-related frailty progression and enhancing quality of life. Measures of body composition will become increasingly relevant as evidence for nutritional support, prehabilitation, and rehabilitation grows. 37 , 38 , 39
AI has previously been used to extract body composition metrics from CT scans. A combined model integrating clinical parameters and body composition parameters demonstrated superior predictive capability for disease progression than models based solely on high-risk cytogenetics or Revised International Staging System. 32 However, to our knowledge, this is the first study to extract body composition metrics from WBMRI using AI.
Manual annotation of medical imaging is a time intensive task, requiring specialized knowledge of anatomy. Generally, this task is performed by radiographers or radiologists. For volumetric WBMRI, this task can take up to several hours per scan. Using our automated method speeds up the process to seconds or minutes and frees up valuable operator time.
Furthermore, this study is limited to extracting metrics from WB Dixon MRI, and we have yet to exploit the quantitative metrics from diffusion-weighted MRI (apparent diffusion coefficient that is inversely related to cellularity), which is also a core component of contemporary WBMRI protocols in oncology.
This study is exploratory and single center with modest sample sizes for models involving later time points; model selection across many candidate predictors risks overfitting despite using BIC to encourage parsimony. Only linear relationships between candidate predictors and PFS were investigated, which does not accommodate potential nonlinear associations. Imaging was acquired on a single scanner and the pipeline required retraining for disease specific features, so external validation is essential. We did not incorporate functional measures (strength, performance status) or detailed metabolic biomarkers, limiting causal inference about mechanisms. Replication in independent, multicenter cohorts and prospective validation of prespecified IDP-based prognostic models are now priorities. Furthermore, comparison with established measures of frailty such as the International Myeloma Working Group frailty score should be planned in new studies. 40
We have demonstrated that a retrained deep-learning pipeline can reliably extract automated IDPs from routinely acquired clinical WBMRI in patients with MM, and that baseline and longitudinal IDPs show consistent associations with PFS, even when accounting for conventional clinical metrics. These findings support opportunistic body composition phenotyping from diagnostic WBMRI as a scalable biomarker for risk stratification and mechanistic study.
Conflict-of-interest disclosure: N.B. and B.W. are cofounders of Orcino Health Ltd. There is no financial relationship between the Royal Marsden Hospital and Orcino Health Ltd. The remaining authors declare no competing financial interests.
References
1 Hillengass J Usmani S Rajkumar SV International Myeloma Working Group consensus recommendations on imaging in monoclonal plasma cell disorders Lancet Oncol 20 6 2019 e302 e312 31162104 10.1016/S1470-2045(19)30309-2
2 Myeloma: Diagnosis and Management 2016 National Institute for Health Care and Excellence https://www.nice.org.uk/guidance/ng35 26913316
3 Kaiser M Goh V Stern S BSH Committee Advanced imaging for earlier diagnosis and morbidity prevention in multiple myeloma: a British Society of Haematology and UK Myeloma Society Good Practice Paper Br J Haematol 205 4 2024 1319 1325 39248274 10.1111/bjh.19716
4 Messiou C Porta N Koh DM Whole body MRI by MY-RADS for imaging response assessment in multiple myeloma Blood Cancer J 15 1 2025 122 40675973 10.1038/s41408-025-01327-4 PMC12271311
5 Belotti A Ribolla R Cancelli V Predictive role of diffusion-weighted whole-body MRI (DW-MRI) imaging response according to MY-RADS criteria after autologous stem cell transplantation in patients with multiple myeloma and combined evaluation with MRD assessment by flow cytometry Cancer Med 10 17 2021 5859 5865 34263564 10.1002/cam4.4136 PMC8419770
6 Messiou C Hillengass J Delorme S Guidelines for acquisition, interpretation, and reporting of whole-body MRI in myeloma: myeloma response assessment and diagnosis system (MY-RADS) Radiology 291 1 2019 5 13 30806604 10.1148/radiol.2019181949
7 Padhani AR Lecouvet FE Tunariu N METastasis reporting and data system for prostate cancer: practical guidelines for acquisition, interpretation, and reporting of whole-body magnetic resonance imaging-based evaluations of multiorgan involvement in advanced prostate cancer Eur Urol 71 1 2017 81 92 27317091 10.1016/j.eururo.2016.05.033 PMC5176005
8 Petralia G Koh DM Attariwala R Oncologically relevant findings reporting and data system (ONCO-RADS): guidelines for the acquisition, interpretation, and reporting of whole-body MRI for cancer screening Radiology 299 2021 494 507 33904776 10.1148/radiol.2021201740
9 Cook G Larocca A Facon T Zweegman S Engelhardt M Defining the vulnerable patient with myeloma—a frailty position paper of the European Myeloma Network Leukemia 34 9 2020 2285 2294 32555295 10.1038/s41375-020-0918-6 PMC7449877
10 Cruz-Jentoft AJ Bahat G Bauer J Writing Group for the European Working Group on Sarcopenia in Older People 2 EWGSOP2, and the Extended Group for EWGSOP2 Sarcopenia: revised European consensus on definition and diagnosis Age Ageing 48 4 2019 601 631 10.1093/ageing/afz046 PMC6593317 31081853
11 Nandakumar B Baffour F Abdallah NH Sarcopenia identified by computed tomography imaging using a deep learning–based segmentation approach impacts survival in patients with newly diagnosed multiple myeloma Cancer 129 3 2023 385 392 36413412 10.1002/cncr.34545 PMC9822865
12 Yip C Dinkel C Mahajan A Siddique M Cook GJ Goh V Imaging body composition in cancer patients: visceral obesity, sarcopenia and sarcopenic obesity may impact on clinical outcome Insights Imaging 6 4 2015 489 497 26070723 10.1007/s13244-015-0414-0 PMC4519815
13 Shah UA Whiting K Devlin S Extreme body mass index and survival in newly diagnosed multiple myeloma patients Blood Cancer J 13 1 2023 13 36631444 10.1038/s41408-022-00782-7 PMC9834289
14 Liu Y Basty N Whitcher B Genetic architecture of 11 organ traits derived from abdominal MRI using deep learning elife 10 2021 e65554 10.7554/eLife.65554 PMC8205492 34128465
15 Basty N Sorokin EP Thanaj M Abdominal imaging associates body composition with COVID-19 severity PLoS one 18 4 2023 13 e0283506 10.1371/journal.pone.0283506 PMC10101472 37053189
16 Thanaj M Basty N Whitcher B Precision MRI phenotyping of muscle volume and quality at a population scale Front Physiol 15 2024 1288657 10.3389/fphys.2024.1288657 PMC10869600 38370011
17 Rata M Blackledge M Scurr E Implementation of whole-body MRI (MY-RADS) within the OPTIMUM/MUKnine multi-centre clinical trial for patients with myeloma Insights Imaging 13 1 2022 123 35900614 10.1186/s13244-022-01253-0 PMC9334517
18 Basty N Liu Y Cule M Thomas EL Bell JD Whitcher B Image processing and quality control for abdominal magnetic resonance imaging in the UK Biobank arXiv Preprint posted online 16 July 2020 10.48550/arXiv.2007.01251
19 Harada T Abe T Kato F Five-point Likert scaling on MRI predicts clinically significant prostate carcinoma BMC Urol 15 1 2015 91 26334874 10.1186/s12894-015-0087-5 PMC4558639
20 Messiou C Porta N Sharma B Prospective evaluation of whole-body MRI versus FDG PET/CT for lesion detection in participants with myeloma Radiol Imaging Cancer 3 5 2021 e210048 10.1148/rycan.2021210048 PMC8489453 34559006
21 Kumar S Paiva B Anderson KC International Myeloma Working Group consensus criteria for response and minimal residual disease assessment in multiple myeloma Lancet Oncol 17 8 2016 e328 e346 27511158 10.1016/S1470-2045(16)30206-6
22 Durie BGM Harousseau J-L Miguel JS International Myeloma Working Group International uniform response criteria for multiple myeloma Leukemia 20 9 2006 1467 1473 16855634 10.1038/sj.leu.2404284
23 Rajkumar SV Dimopoulos MA Palumbo A International Myeloma Working Group updated criteria for the diagnosis of multiple myeloma Lancet Oncol 15 12 2014 e538 e548 25439696 10.1016/S1470-2045(14)70442-5
24 Schwarz G Estimating the dimension of a model Ann Statist 6 2 1978 461 464
25 Harrell FE Califf RM Pryor DB Lee KL Rosati RA Evaluating the yield of medical tests Jama 247 18 1982 2543 2546 7069920
26 Stege CA Nasserinejad K Levin MD Geriatric impairments and low muscle mass are associated with treatment discontinuation and overall survival in newly diagnosed non-transplant eligible multiple myeloma patients (nte-NDMM) treated with dose-adjusted melphalan-prednisone-bortezomib (MPV)—results of the Dutch HOVON 123 study Blood 132 suppl 1 2018 1889
27 Abdallah NH Nagayama H Takahashi N Muscle and fat composition in patients with newly diagnosed multiple myeloma Blood Cancer J 13 1 2023 185 38086801 10.1038/s41408-023-00934-3 PMC10716405
28 Williams A Baruah D Patel J Prevalence and significance of sarcopenia in multiple myeloma patients undergoing autologous hematopoietic cell transplantation Bone Marrow Transpl 56 1 2021 225 231 10.1038/s41409-020-01008-9 PMC7855279 32732941
29 Takeoka Y Sakatoku K Miura A Prognostic effect of low subcutaneous adipose tissue on survival outcome in patients with multiple myeloma Clin Lymphoma Myeloma Leuk 16 8 2016 434 441 27263047 10.1016/j.clml.2016.04.010
30 Bhamidipati PK Carson KR Wildes TM Psoas cross-sectional area as radiographic measure of sarcopenia does not predict overall survival in multiple myeloma Blood 122 21 2013 5326
31 Lee B Han HS Yoon YS Park Y Kang M Kim J “Obesity Paradox” as a new insight from long-term survivors in pancreatic cancer patients HPB (Oxford) 27 7 2025 922 929 40189966 10.1016/j.hpb.2025.03.450
32 Wegner F Sieren MM Grasshoff H AI-based body composition analysis of CT data has the potential to predict disease course in patients with multiple myeloma Sci Rep 15 1 2025 26455 10.1038/s41598-025-11560-3 PMC12280154 40691492
33 Prado CM Gonzalez MC Heymsfield SB Body composition phenotypes and obesity paradox Curr Opin Clin Nutr Metab Care 18 6 2015 535 551 26335310 10.1097/MCO.0000000000000216
34 Groß JP Nattenmüller J Hemmer S Body fat composition as predictive factor for treatment response in patients with newly diagnosed multiple myeloma–subgroup analysis of the prospective GMMG MM5 trial Oncotarget 8 40 2017 68460 10.18632/oncotarget.19536 PMC5620270 28978130
35 Kylies J Brauneck E Priemel M Sarcopenia and fat loss from serial CT predict survival in multiple myeloma patients undergoing stem cell transplantation World J Surg Oncol 23 1 2025 336 40963145 10.1186/s12957-025-04007-6 PMC12442264
36 Do TD Nonnenmacher T Burghardt M AI-based 3D-segmentation quantifies sarcopenia in multiple myeloma patients Diagnostics 15 19 2025 2466 41095685 10.3390/diagnostics15192466 PMC12523729
37 Lee K Nathwani N Shamunee J Telehealth exercise to improve physical function and frailty in patients with multiple myeloma treated with autologous hematopoietic Stem cell transplantation (TIPS): protocol of a randomized controlled trial Trials 23 1 2022 921 36329525 10.1186/s13063-022-06848-y PMC9633031
38 Normann AJ Mo CC Wilson RL Prehabilitation exercise training to target improved muscle strength in pretransplant patients diagnosed with multiple myeloma: protocol for a pilot randomized controlled trial JMIR Res Protoc 13 1 2024 e64905 10.2196/64905 PMC11695955 39701583
39 Land J Luong MK Longden A Real-world evaluation of physiotherapist-led exercise prehabilitation and rehabilitation during autologous stem cell transplantation in myeloma: a single-centre experience BMJ Open Qual 14 1 2025 e002936 10.1136/bmjoq-2024-002936 PMC11752014 39762061
40 Palumbo A Bringhen S Mateos MV Geriatric assessment predicts survival and toxicities in elderly myeloma patients: an International Myeloma Working Group report Blood 125 13 2015 2068 2074 25628469 10.1182/blood-2014-12-615187 PMC4375104
Supplementary Material
Supplemental Methods, Tables and Figure Supplemental Tables
Acknowledgments
The authors thank all participating investigators, centers, and participating patients and their families.
This work represents independent research funded by the National Institute for Health and Care Research ( 10.13039/100014461 NIHR) Biomedical Research Centre at The Royal Marsden NHS Foundation Trust and The Institute of Cancer Research , London, United Kingdom, and by the Royal Marsden Cancer Charity and the CRUK National Cancer Imaging Translational Accelerator .
The views expressed are those of the authors and not necessarily those of the NIHR or the Department of Health and Social Care.
Authorship
Contribution: C.M., B.W., and N.B. designed the study; N.B. and B.W. performed image analysis; B.W. analyzed the data with oversight from K.T.; C.M., B.W., and N.B. wrote the manuscript; M.K., C.P., K.B., K.S., K.T., N.P., D.M., R.E., J.E., A.D., S.D., and D.-M.K. were involved in data collection and curation; and all authors revised and approved the manuscript.
The data sets generated and/or analyzed during this study, as well as the multiple myeloma processing and segmentation pipeline, are available from the corresponding author, Christina Messiou ( christina.messiou@rmh.nhs.uk ), on reasonable request.
The full-text version of this article contains a data supplement.