Deep learning and transformer-based feature fusion of conventional MRI for differentiating spinal osteolytic bone metastases and multiple myeloma.

OBJECTIVE

To develop a deep learning model utilizing conventional MRI sequences integrated with Transformer-based feature fusion to differentiate spinal osteolytic bone metastases (OBM) from multiple myeloma (MM).

MATERIALS AND

METHODS

This retrospective study included 663 patients (mean age: 62.13 ± 9.57 years; 378 males) from two medical centers, comprising 342 cases of OBM and 321 cases of MM. All patients underwent MRI examinations with T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and fat-suppressed T2-weighted imaging (T2WI-fs) sequences. Deep learning features were extracted with DenseNet169 to construct classification models, including Multilayer Perceptron (MLP), Support Vector Machine (SVM), Gradient Boosting, AdaBoost, and Naive Bayes. Both feature-level fusion and Transformer-based fusion methods were applied to enhance diagnostic performance. Models were trained (n = 421) and externally tested (n = 242). Their performance was evaluated using receiver operating characteristic (ROC) curves, calibration plots, and decision curve analysis (DCA), with the AUC serving as the primary evaluation metric. The Delong test was used to compare model performance.

RESULTS

Among the single-sequence models, T2WI-fs achieved the highest performance, with an AUC of 0.765 and an accuracy of 0.726. Among the fusion methods, the T2WI+T2WI-fs_Transformer fusion model performed best, with an AUC of 0.783, an accuracy of 0.723, followed by the T2WI+T2WI-fs_early fusion model (AUC = 0.762). Although the differences among models were not statistically significant in the external test set (all P > 0.05), the Transformer fusion model demonstrated superior clinical net benefit and robust generalizability.

CONCLUSION

Transformer-based feature fusion of conventional MRI sequences enables accurate, non-invasive differentiation between spinal OBM and MM, providing significant clinical utility.