AI-derived whole-body MRI metrics in multiple myeloma: treatment-related body composition change and its association with outcomes

A whole-body MRI segmentation pipeline built on a healthy population (UK Biobank) was transferred to myeloma patients and, after fine-tuning on 40 routine clinical scans (20 for training, 19 for validation, one discarded for a DICOM technical error), exceeded the predefined Dice thresholds for every structure, with a mean Dice of 0.916 and mean Likert of 4.58. The authors state this is the first study to extract body composition metrics from whole-body MRI (rather than CT) using AI, showing that models trained on population cohorts can be adapted to clinical images containing disease-specific features such as background hypercellular marrow, focal marrow lesions and metal prosthesis artifacts. Success was defined by per-structure Dice thresholds (0.9 for body composition structures, 0.85 for internal organs) and independently rated by a consultant radiologist and a senior research radiographer on a 5-point Likert scale, averaged across structures and assessors; the sample was 40 single-center scans.

In the iTIMM cohort, most image-derived phenotypes changed measurably during treatment: abdominal skeletal muscle volume decreased from MRI1 to MRI2 (P < .001) and then did not change; subcutaneous and visceral adipose tissue increased from MRI1 to MRI2 and decreased from MRI2 to MRI3 (both P < .001); liver relative fat fraction increased from MRI1 to MRI2 (P < .005) and overall from MRI1 to MRI3 (P < .001); heart volume decreased from MRI1 to MRI2 (P < .001); kidney volumes decreased from MRI2 to MRI3 (P = .029 and P = .023); and both iliopsoas muscles decreased from MRI1 to MRI3 (both P < .001). This extends whole-body MRI beyond lesion detection and response assessment toward quantitative follow-up of non-diseased tissue without additional scanning burden; the authors note that cardiac segmentation had the lowest Dice scores, possibly due to cardiac motion and variable inclusion of the aorta and vena cava, so the clinical significance of that finding is uncertain. Prospective single-center observational cohort with 69 evaluable patients (67 for MRI3 analyses, two with corrupted imaging), median follow-up of 42 months; Pearson correlation for same-scan associations and paired t tests for between-scan differences.

Across six modeling scenarios, higher baseline (MRI1) abdominal skeletal muscle and subcutaneous adipose tissue were associated with longer progression-free survival in all three models; increasing visceral adipose tissue over time was associated with shorter progression-free survival in all three models; increasing liver relative fat fraction over time was associated with longer progression-free survival in all three models; and higher baseline abdominal skeletal muscle was associated with longer progression-free survival in two of three models. The authors report these associations persisted after accounting for conventional clinical metrics, and note that BMI as a crude measure may mask divergent effects of lean and fat compartments: binarized abdominal skeletal muscle and subcutaneous adipose tissue (thresholds 2.63 L/m² and 1.34 L/m²) showed clear separation in Kaplan-Meier curves, whereas binarized BMI (26.7 kg/m²) did not. Cox proportional hazards models with 45 progression-free survival events (overall survival had 13 events and was not the primary endpoint); three 'best' models were selected from all possible variable combinations by balancing Bayesian information criterion and concordance index, and the authors note selection was not stepwise, so variables recurring across models with similar performance indicate a more robust association.

The authors propose that body composition changes in myeloma are likely strongly influenced by steroid administration, a cornerstone of induction therapy that increases appetite and leads to increases in visceral adipose tissue, subcutaneous adipose tissue and liver fat fraction and decreases in skeletal muscle, so baseline liver fat fraction and its rise over time may relate to steroid therapy effects. This offers a treatment-mechanism-related interpretive frame for the observed fat gain and muscle loss, and notes that the association of rising visceral adipose tissue with poorer progression-free survival is concordant with CT-derived reports linking visceral adipose tissue to inferior response to bortezomib-based induction therapy. This is a discussion-level mechanistic interpretation and literature comparison; the study did not incorporate functional measures (strength, performance status) or detailed metabolic biomarkers, which the authors state limits causal inference about mechanisms.

The pipeline applies to T1-weighted Dixon whole-body MRI data acquired with the Myeloma Response Assessment and Diagnosis System consensus protocol from head to knees, and the intended population is adults with newly diagnosed or first-relapsed symptomatic myeloma planned for induction therapy followed by high-dose melphalan and autologous stem cell transplant. The authors argue that standardized acquisition, including breath-holds to reduce breathing motion, supports translation to data acquired at other sites and in other diseases, and that automation compresses manual contouring from up to several hours per scan to seconds or minutes, freeing radiographer and radiologist time. They also note the study has not yet exploited quantitative metrics from diffusion-weighted MRI (apparent diffusion coefficient, inversely related to cellularity), which is a core component of contemporary oncological whole-body MRI protocols.

Open questions the authors raise include that the study is exploratory and single center with modest sample sizes for models involving later time points, that model selection across many candidate predictors risks overfitting despite using the Bayesian information criterion to encourage parsimony, that only linear relationships between candidate predictors and progression-free survival were investigated so potential nonlinear associations are not accommodated, that imaging was acquired on a single scanner and the pipeline required retraining for disease-specific features so external validation is essential, and that functional measures (strength, performance status) and detailed metabolic biomarkers were not incorporated, limiting causal inference about mechanisms. They list replication in independent multicenter cohorts, prospective validation of prespecified image-derived-phenotype-based prognostic models, and comparison with established frailty measures such as the International Myeloma Working Group frailty score as priorities. In addition, cardiac segmentation had the lowest Dice scores, possibly due to cardiac motion and variable inclusion of the aorta and vena cava, so the clinical significance of that finding remains uncertain; this is a full-text parse in which the specific values in Figures 3 and 4 and supplemental Tables 1-7 and supplemental Figures 1-2 are not given in the body text, so exact effect sizes would require consulting the original supplementary material.