Benign and malignant bone lesion diagnosis based on self-supervised and radiomics fusion using SPECT/CT images.

Purpose: This study aims to improve the diagnostic accuracy of SPECT/CT imaging in distinguishing benign from malignant bone lesions by integrating self-supervised deep learning and radiomics, reducing subjectivity in traditional image interpretation for more reliable clinical decision-making. Methods: We developed a multi-scale, multi-modal framework combining radiomics with self-supervised learning. The novel SPECT-guided model, SPARC-Net, uses functional SPECT data as semantic priors to extract discriminative features from CT scans without manual annotations. Deep features from SPARC-Net were fused with radiomics to form a unified representation. The model was trained and validated on 741 confirmed bone lesion cases using five-fold cross-validation, with interpretability assessed via Grad-CAM. Results: The fused model achieved 82.3 % accuracy, 0.890 AUC, 72.3 % F1 score, 79.3 % precision, 66.6 % sensitivity, and 90.7 % specificity, outperforming single-modality models. Grad-CAM confirmed the model focused on metabolically active regions identified by SPECT. Conclusion: SPARC-Net, integrating SPECT-guided self-supervised learning with CT-based radiomics, improves classification of benign and malignant lesions, enhancing the accuracy, robustness, and interpretability of SPECT/CT imaging for bone tumor diagnosis.