A semantic-enhanced integrated radiomics model based on spinal MRI for predicting early relapse in multiple myeloma: A multi-center study

Motivation: Early relapse (ER) in MM leads to poor prognosis, making early identification of high-risk patients essential. Current models are limited by biopsy variability and high gene detection costs. A non-invasive ER prediction tool for MM is urgently needed. Goal(s): Develop and validate a radiomics model combining MRI, clinical, and semantic features to predict ER in MM. Approach: We used multicenter retrospective data and 5-fold cross-validation, evaluating performance on an external test set with ROC and decision curves. Results: The model (T2 + SF + clinic) showed AUC of 0.856, good calibration, and high net benefit. ER patients had shorter PFS (P < 0.001). Impact: Integrating imaging and semantic features improved model accuracy and interpretability, enabling early risk identification and supporting personalized MM treatment.