Abstract
We investigated whether radiomics of computed tomography (CT) image data enables the differentiation of bone metastases not visible on CT from unaffected bone, using pathologically confirmed bone metastasis as the reference standard, in patients with gastric cancer. In this retrospective study, 96 patients (mean age, 58.4 ± 13.3 years; range, 28-85 years) with pathologically confirmed bone metastasis in iliac bones were included. The dataset was categorized into three feature sets: (1) mean and standard deviation values of attenuation in the region of interest (ROI), (2) radiomic features extracted from the same ROI, and (3) combined features of (1) and (2). Five machine learning models were developed and evaluated using these feature sets, and their predictive performance was assessed. The predictive performance of the best-performing model in the test set (based on the area under the curve [AUC] value) was validated in the external validation group. A Random Forest classifier applied to the combined radiomics and attenuation dataset achieved the highest performance in predicting bone marrow metastasis in patients with gastric cancer (AUC, 0.96), outperforming models using only radiomics or attenuation datasets. Even in the pathology-positive CT-negative group, the model demonstrated the best performance (AUC, 0.93). The model's performance was validated both internally and with an external validation cohort, consistently demonstrating excellent predictive accuracy. Radiomic features derived from CT images can serve as effective imaging biomarkers for predicting bone marrow metastasis in patients with gastric cancer. These findings indicate promising potential for their clinical utility in diagnosing and predicting bone marrow metastasis through routine evaluation of abdominopelvic CT images during follow-up.