The SPECT/CT radiomics-based classification of skeletal metastases and benign bone lesions.
Authors
Affiliations (7)
Affiliations (7)
- Department of Computational Medicine, Ewha Womans University, College of Medicine, Seoul, Korea.
- Department of Nuclear Medicine, Ewha Womans University, College of Medicine, Seoul, Korea.
- Department of Biomedical Engineering, Ewha Womans University, College of Medicine, Seoul, Korea.
- Ewha Medical Research Institute Ewha Womans University, College of Medicine, Seoul, Korea.
- Ewha Medical Artificial Intelligence Research Institute Ewha Womans University, College of Medicine, Seoul, Korea.
- Department of Computational Medicine, Ewha Womans University, College of Medicine, Seoul, Korea. [email protected].
- Department of Nuclear Medicine, Ewha Womans University, College of Medicine, Seoul, Korea. [email protected].
Abstract
Hybrid Single Photon Emission Computed Tomography/Computed Tomography (SPECT/CT) improves lesion localization and diagnostic accuracy in detecting skeletal metastases. However, interpretation based on standardized uptake value (SUV) remains suboptimal due to overlapping uptake between benign and malignant lesions. Radiomics quantifies intralesional heterogeneity and may overcome SUV limitations. This study aimed to evaluate whether integrating SPECT/CT radiomics with SUV improves differentiation between skeletal metastases and benign bone lesions. Sixty-three patients who underwent SPECT/CT between September 2021 and March 2024 at Ewha Mokdong Hospital were included. Benign and metastatic lesions were classified based on histopathology or imaging follow-up. Four classification models were developed based on features: (1) SUV<sub>max</sub>, (2) SPECT radiomics, (3) SUV<sub>max</sub> with SPECT radiomics (SUV<sub>max</sub> + SPECT), and (4) SUV<sub>max</sub> with SPECT and CT radiomics (SUV<sub>max</sub> + SPECT/CT). Radiomic features were extracted using PyRadiomics, followed by feature selection via Least absolute shrinkage and selection operator (LASSO), Bhattacharyya distance and Boruta algorithm. Machine learning classifiers included Support vector machines (SVM), Extreme gradient boosting (XGBoost), and Random Forest. Models were evaluated on a test set. For the SUV<sub>max</sub> model, both the SVM and Random Forest classifiers achieved the same area under the curve (AUC) scores of 0.89 on the test set. In the SPECT radiomics model, the SVM with LASSO exhibited an AUC of 0.90. SUV<sub>max</sub> + SPECT with Bhattacharyya-XGBoost achieved an AUC of 0.92. The highest performance (AUC 0.93) was achieved by SUV<sub>max</sub> + SPECT/CT with Boruta-SVM. SPECT and CT radiomics improved diagnostic differentiation beyond SUV<sub>max</sub> alone. SPECT radiomics demonstrated strong standalone potential, while integration with CT radiomics further enhanced performance. These findings support radiomics as a complementary tool for more accurate and individualized lesion classification in nuclear medicine.