Lesion-level dual-tracer PET biomarkers predict prognosis in multiple myeloma treated with CXCR4-directed radiopharmaceutical therapy

Background C-X-C motif chemokine receptor 4 (CXCR4)-directed radiopharmaceutical therapy (RPT) represents a promising option for hematologic malignancies. Nevertheless, responses in relapsed and refractory (r/r) multiple myeloma (MM) are heterogenous, emphasizing the need for optimized patient selection before RPT. Current approaches mostly rely on qualitative assessments, confirming relevant CXCR4-expression by CXCR4-directed PET in vital myeloma burden, the latter usually being determined by additional [ 18 F]FDG-PET. Purpose To evaluate whether quantitative imaging biomarkers derived from dual-tracer PET/CT can enhance patient stratification and predict therapeutic response and survival following CXCR4-RPT. Materials and methods 22 patients with r/r MM who underwent CXCR4-directed [⁶⁸Ga]Ga-PentixaFor-PET/CT and [¹⁸F]FDG-PET/CT imaging prior to CXCR4-RPT were retrospectively analyzed. A fully automated pipeline performed deep-learning–based lesion segmentation and dual-tracer fusion; batch quality control and correction ensured segmentation accuracy and concordant-lesion adjudication (> 10% volumetric overlap) prior to lesion-level feature extraction. Features included demographics, laboratory/genetic variables, and imaging metrics from FDG- and CXCR4-PET, stratified by anatomic site (medullary vs. extramedullary) and concordance (concordant vs. discordant as defined by comparing FDG- and CXCR4-positive myeloma lesions). Surface-standardized maximum inter-lesion distances (sDmax) were additionally computed. Endpoints were therapy response (responder vs. non-responder as defined by serological response assessment based on IMWG-criteria or PET/MRI based response assessment) and overall survival (OS). Group comparisons were performed using Welch’s t-test/Chi-square; survival analysis was conducted applying Kaplan–Meier estimates and log-rank tests. Decision-tree models were interpreted with SHapley Additive exPlanations (SHAP). Results Responders to RPT showed lower [ 18 F]FDG-SUV mean in medullary and extramedullary concordant lesions (Welch’s p = 0.03). In the response classifier, these metrics ranked among the top predictors by SHAP, alongside selected extramedullary CXCR4-dominant discordant features and high-risk cytogenetics. Shorter OS was associated with higher TLG[FDG medullary concordant], greater sDmax[FDG], and higher CXCR4-positive medullary burden (MTV/TLC) (log-rank p < 0.05). SHAP directionality agreed with univariate trends. BMI showed a modest inverse association with early mortality in the 6-month survival model. Conclusions In our exploratory analysis, [¹⁸F]FDG-uptake within medullary concordant lesions was linked to both response and survival, while CXCR4-expressing medullary volume and sDmax added survival information. Concordance-aware, lesion-level quantification from dual-tracer PET/CT might present as a promising approach to aid risk stratification for CXCR4-directed RPT. However, initial findings of our hypothesis generating analysis warrant prospective validation in larger cohorts.