First report of AI derived metrics from whole-body MRI in patients with multiple myeloma reveals unique insights into body composition and changes on therapy

Abstract Background: Data on sensitivity and cost effectiveness has shaped national and International Myeloma Working Group guidelines to include non-invasive whole-body MRI (WBMRI) for Multiple Myeloma (MM) diagnosis and follow up. Furthermore, acquisition has been standardised (MY-RADS) in line with other oncological WBMRI protocols. This presents an opportunity to complement accurate disease detection and response assessments with body composition metrics beyond traditional measures such as BMI and sarcopaenia. The aim of this study was to define the technical capability of AI to derive quantitative image-derived phenotypes (IDPs) in non-diseased tissue from WBMRI in patients with MM and to derive longitudinal measurements and exploratory associations with imaging response. Methods: An AI-based segmentation algorithm was trained to quantify the volume of skeletal muscle, abdominal subcutaneous adipose tissue (ASAT) and visceral adipose tissue (VAT), internal organs, bones and relative fat fraction (rFF) of the liver. The model (originally trained on >100 UK Biobank WBMRIs) was optimised for patients with MM using WBMRI scans from 20 MM patients. The resulting pipeline was applied to WBMRI scans from 69 MM patients with active MM who had been recruited to the iTiMM study (NCT02403102) and imaged at 3 time points: (1) baseline, (2) post-induction chemotherapy, and (3) post-autologous stem cell transplant (ASCT). Changes in IDPs between time points were performed using paired univariate hypothesis tests. MY-RADS response assessment categories (RAC score 1 (best response) vs 2 or greater), were used to evaluate associations between IDPs and radiological response. Volumetric IDPs were divided by height squared to minimize variability due to patient size. Associations between IDPs and MY-RADS RAC score were measured with logistic regression models, selected using a penalized model selection criterion, that included both clinical measurements and IDPs. Results: The segmentation model achieved very high performance, with the average Dice score across all structures being 0.916 [range 0.812-0.99]. Longitudinal analysis revealed changes in normalised IDPs following therapy, where skeletal muscle and cardiac volume decreased and ASAT and VAT increased between the baseline WBMRI and the post-induction WBMRI. ASAT and VAT then decreased between the post-induction WBMRI and the post-ASCT WBMRI falling to below baseline. Liver rFF increased between the baseline WBMRI and the post-induction WBMRI but did not change thereafter. BMI increased on average between the baseline and post-induction WBMRI and decreased between the post-induction WBMRI and the post-ASCT WBMRI. A regression model using baseline age, total disease burden score, skeletal muscle, ASAT, VAT, and kidney volume was associated with MY-RADS RAC score on the post-ASCT WBMRI. More VAT at baseline was associated with an excellent MY-RADS RAC score of 1 post-ASCT. Increased skeletal muscle, ASAT, and age at baseline were associated with an inferior MY-RADS RAC score of 2 or greater. Conclusions To our knowledge this is the first study of AI derived body composition metrics using standard WBMRI data in patients with multiple myeloma. The AI-based IDP pipeline has been successfully trained for WBMRI in MM patients, overcoming the challenges posed by skeletal deformities and abnormal bone marrow signal. This study demonstrated significant tissue-specific changes over time, including a reduction in skeletal muscle and increases in ASAT, VAT and liver rFF following induction chemotherapy which we note correspond with the period of high dose steroid therapy. Higher VAT at baseline was also associated with best imaging response on the post ASCT WBMRI. These associations suggest potential links between body composition dynamics and radiological treatment response and could be in line with the obesity paradox, where greater adiposity is sometimes associated with improved outcome. However, weight dependent dosing of bortezomib and melphalan may also contribute. The ability to apply AI to routine WBMRI scans offers an objective and scalable approach to monitor tissue-specific changes during treatment and to amplify the impact of AI-derived metrics of disease burden and skeletal deformity also under development. This study highlights the potential of IDPs and their role towards a more holistic view of the patient and non invasive insights into body composition, wellness and toxicity.