Radiomics Models for the Preoperative Prediction of Pelvic and Sacral Tumor Types: A Single-Center Retrospective Study of 795 Cases

PURPOSE: To assess the performance of random forest (RF)-based radiomics approaches based on 3D computed tomography (CT) and clinical features to predict the types of pelvic and sacral tumors. MATERIALS AND METHODS: A total of 795 patients with pathologically confirmed pelvic and sacral tumors were analyzed, including metastatic tumors (n = 181), chordomas (n = 85), giant cell tumors (n =120), chondrosarcoma (n = 127), osteosarcoma (n = 106), neurogenic tumors (n = 95), and Ewing’s sarcoma (n = 81). After semi-automatic segmentation, 1316 hand-crafted radiomics features of each patient were extracted. Four radiomics models (RMs) and four clinical-RMs were built to identify these seven types of tumors. The area under the receiver operating characteristic curve (AUC) and accuracy (ACC) were used to evaluate different models. RESULTS: In total, 795 patients (432 males, 363 females; mean age of 42.1 ± 17.8 years) were consisted of 215 benign tumors and 580 malignant tumors. The sex, age, history of malignancy and tumor location had significant differences between benign and malignant tumors (P < 0.05). For the two-class models, clinical-RM2 (AUC = 0.928, ACC = 0.877) performed better than clinical-RM1 (AUC = 0.899, ACC = 0.854). For the three-class models, the proposed clinical-RM3 achieved AUCs between 0.923 (for chordoma) and 0.964 (for sarcoma), while the AUCs of the clinical-RM4 ranged from 0.799 (for osteosarcoma) to 0.869 (for chondrosarcoma) in the validation set. CONCLUSIONS: The RF-based clinical-radiomics models provided high discriminatory performance in predicting pelvic and sacral tumor types, which could be used for clinical decision-making.