The deep learning radiomics nomogram for risk stratification in multiple myeloma using automatic whole-body [¹⁸F]FDG PET/CT segmentation approach.

Abstract

To develop a deep learning (DL) approach for automatic segmentation and accurate risk stratification in multiple myeloma (MM) using whole-body [¹⁸F]FDG PET/CT. This retrospective study included MM patients who underwent [¹⁸F]FDG PET/CT between August 2013 and December 2023. To automatically segment focal lesions, an nnU-Net architecture was trained using dual-channel PET/CT inputs. For diffuse/mixed patterns, the tumor was automatically identified via bone segmentation using liver SUVmedian thresholds. Subsequently, DL and radiomics features were extracted using Pyradiomics and the STU-Net encoder. Then a deep learning radiomics nomogram (DLRN) was constructed using the Cox proportional hazards model and evaluated using the calibration curve, the receiver operating characteristic (ROC) curve, the Kaplan-Meier curve, and decision curve analysis. The study included 345 patients (median age, 59 years [IQR, 35-67 years], 198 male). The nnU-Net achieved a median DSC of 0.64-0.77 for focal lesions segmentation across cohorts. The DLRN was constructed by integrating deep learning radiomics score (DLRS), lactate dehydrogenase (LDH), and β2-microglobulin (β2-MG). The DLRN achieved an area under ROC curve (AUC) of 0.87 (95% confidence interval [CI]: 0.82-0.93), 0.84 (95% CI: 0.73-0.96), and 0.88 (95% CI: 0.76-0.99) for 3-year overall survival (OS) status prediction in the training, internal and external testing cohorts, which outperformed the International Staging System (ISS) (all P < 0.05). Furthermore, the DLRN can effectively identify high-risk individuals (all P < 0.05), demonstrated good agreement between predicted and observed survival probabilities, and provided clinical net benefit. The pattern-specific DL approach achieved automated whole-body tumor segmentation in MM, and the established DLRN demonstrated improved risk stratification capability.