Diagnostic performance of dual-layer spectral CT Radiomics and deep learning for differentiating osteoblastic bone...

This study aimed to compare the diagnostic performance of radiomic features derived from dual-layer spectral detector computed tomography (DLSCT) and a deep learning (DL) model applied to conventional CT images in the differentiation of osteoblastic bone metastases (OBM) from bone islands (BI).

This retrospective study included patients with osteogenic lesions who underwent DLSCT examinations between March 2023 and September 2023. We extracted first-order radiomic features (e.g., mean, maximum, entropy) from both conventional and spectral images. A previously validated DL model was applied to the conventional CT images. We evaluated diagnostic performance using ROC curve analysis, comparing AUC, sensitivity, and specificity.

The study included 216 lesions from 94 patients (66 ± 12 years; 48 males, 46 females): 125 BI and 91 OBM lesions. Significant differences were observed between OBM and BI groups for the mean, maximum, entropy, and uniformity of first-order radiomic features (all P < 0.05). DLSCT (entropy from VMI40keV) and the DL model had comparable AUCs (0.93 vs. 0.96; P = 0.274). However, DLSCT showed superior sensitivity (92 % vs. 62 %; P = 0.002) but comparable specificity (88 % vs. 96 %; P = 0.07) for diagnosing OBM compared to the DL model.

Radiomic features from DLSCT differentiate between BI and OBM with diagnostic performance comparable to that of a DL model. Furthermore, VMI40keV image-derived entropy demonstrated superior sensitivity in diagnosing OBM compared to the DL model.