This retrospective study received approval from the Institutional Ethics Committee, and informed consent from individual subjects was waived. Between January and November 2020, 230 newly diagnosed cancer patients who underwent dual-energy spectral CT were enrolled. The inclusion criteria included: 1) satisfactory image quality without non-negligible artifacts; 2) hyperdense lesions measuring from 3–20 mm; 3) confirmed diagnosis of either BI or OBM; 4) no prior treatment before the spectral CT scan. Those patients were excluded if: 1) primary cancer cannot be determined; 2) the diagnosis of BI or OBM cannot be determined; 3) lesions were larger than 20 mm or smaller than 3 mm; 4) both osteolytic and osteoblastic metastasis were present on CT; 5) there was evidence of compression fracture or pathologic fracture. After applying these criteria, 51 newly diagnosed cancer patients with 204 bone lesions were ultimately selected for the study. The flowchart of patient selection is depicted in Figure 1, and the clinical characteristics of enrolled patients are detailed in Table 1.

In differentiating BI from OBM, we initially assess lesions using MRI, PET-CT, and bone scintigraphy.

1. For lesions preliminarily diagnosed as bone islands: A 12-month follow-up is conducted.

1) If the lesion remains stable, it is confirmed as a bone island. 2) If interval growth is observed, it is reclassified as an osteoblastic metastasis.

2. For lesions initially suspected to be osteoblastic metastases:

The research employed a 256-row multidetector CT scanner (Revolution CT 750 HD, GE Healthcare, Waukesha, USA) featuring gemstone spectral imaging (GSI) mode for all examinations. The scanning parameters encompassed a field of view ranging from 40 to 50 cm customized to each patient’s size; a matrix resolution of 512 × 512; voltages of 140 kVp and 80 kVp utilizing an ultra-fast peak kVp switching with 0.5 ms, automated tube current, a helical pitch of 0.992, collimation of 64 ×0.625 mm, rotation time of 0.5 ms, and employing a standard reconstruction kernel. A bolus of contrast agent (Loversol, 350 mg/ml; Jiangsu Hengrui, China) was administered at a standard dosage of 1.5 ml/kg, with a flow rate of 2.0–3.0 ml/s. The venous phase of the contrast-enhanced CT scans was captured 50 seconds after injection. Standard algorithm reconstruction was employed for CT images with a slice thickness of 1.25 mm and an ASIR-V of 50%. The resultant CT images were analyzed by experienced board-certified radiologists (L.Y.Z. and Z.L.), with those unsatisfactory images with a low signal-to-noise ratio or overt motion artifacts excluded.

On the standard DECT workstation (GSI Viewer, AW4.7,GE Healthcare), raw data was reconstructed into 40–140 keV mono-energy images from non-contrast CT and contrast-enhanced CT, and material-decomposition images using two-material decomposition algorithms, including water (calcium) and calcium (water) substrate density images from non-contrast spectral CT and Iodine (calcium) substrate density images from contrast-enhanced spectral CT. Using ITK-SNAP software (version 3.8.0, http://www.itksnap.org), all of the lesions were manually segmented on consecutive non-contrast spectral CT and contrast spectral CT images slice by slice by a radiologist (H.H.L. with 4 years of experience in bone CT imaging), and then reviewed by a senior radiologist (Z.L. with more than 10 years of experience in bone CT imaging). Any discrepancies were dissolved through discussion.

Based on the segmented 3D lesions, radiomics features were extracted using “Pyradiomics” packages for each type of generated image. For each segmented lesion, 14 shape features, 18 histogram features, and 75 texture features were extracted from each of plain single-energy images (40–140 keV), each of enhanced single-energy images (40–140 keV), and each of material-separation images including water (calcium) image, calcium (water) image, and iodine (calcium) image, respectively. Since the shape features extracted were all the same for each type of image, we singled out these features to form an independent shape feature set. Finally, four feature sets were formed, including the shape feature set (n = 14), material-separation feature set (n = [18 + 75] × 3 = 279), plain spectral CT feature set (n = [18 + 75] × 11 = 1023), and enhanced spectral CT feature set (n = [18 + 75] × 11 = 1023).

U-test (P < 0.05) and LASSO analysis were sequentially used to select the most relevant features to the differentiation for each feature set, with the lasso coefficient being used for feature importance ranking.

Random forest was used as the classifier to train the machine learning model. In total, we built five models, including a shape model based on the selected shape feature set, a material-separation model based on the selected material-separation feature set, a plain spectral CT model based on the selected plain spectral CT feature set, an enhanced spectral CT model based on the selected enhanced spectral CT feature set, and a combined model based on the combined feature set. To evaluate the classification performance, we tested the model using cross-validation (10-fold) with sensitivity, specificity, accuracy, and area under the receiver operating characteristic curve (AUC) collected.

All continuous variables were denoted as mean ± standard deviation. On SPSS 26.0 (SPSS Incorporation, Chicago, IL), an independent sample student-t test was used to test the difference in demographic data, including age between the BI group and the OBM group. DeLong test (10) was used to compare the performance between each two of the shape model, material-separation model, plain spectral CT model, enhanced spectral CT model, and combined model using MedCalc (version 19.0.4; MedCalc Software). Decision curve analysis (DCA) was applied to quantify the model’s net clinical benefit using the R statistical package (MathSoft Inc., Seattle, WA). Two-tailed P < 0.05 indicates a statistically significant difference.

This study included 51 newly diagnosed cancer patients with 204 bone lesions (116 OBM and 88 BIs). The most common location is the vertebrae both for the BI group (25/88) and the OBM group (73/116). For primary cancer, OBM were predominantly originating from lung cancer (10/14) in our study. There were no significant differences in age and gender between the two groups. However, lesions in the OBM group were significantly larger than those in the BI group (P < 0.001). The patient characteristics are shown in Table 1.

Of the 14 shape features, 13 features were selected after U-test and 4 features were selected after LASSO analysis (Figure 2). For the material-separation feature set with a total of 279 features, 197 features were selected after U-test, and 7 features were selected after LASSO analysis, including 5 features from calcium (water) image and 2 features from iodine (calcium) image (Figure 2).

A total of 1023 features were extracted from the plain spectral CT. 644 features were selected after U-test, and 7 features were selected after LASSO analysis in the plain spectral CT feature set (Figure 2). These selected features included 3 features extracted from 40 keV CT image, 1 feature from 70 keV CT image, 1 feature from 100 keV CT image, and 2 features from 140 keV CT image. The selected plain single-energy CT features showed a trend of having more features from low and high single-energy images (3 features from 40 keV CT image and 2 features from 140 keV CT image) and fewer from intermediate single-energy images (1 feature from 70 keV and 1 feature of 100 keV (Figures 3, 4).

For the 1023 features in the enhanced feature set, 643 features were selected after U-test, and 9 features were selected after LASSO analysis (Figure 2). These 9 features included 3 features from 40 keV CT image, 1 features from 50 keV CT image, 2 features from 60 keV CT image, 1 feature from 80 keV CT image, 1 feature from 120 keV CT image, and 1 feature from 140 keV CT image. However, the selected enhanced single-energy CT features exhibited a trend of having more low-energy features (3 features from 40 keV CT image) and fewer intermediate to high-energy images (1 features from 50 keV CT image, 2 features from 60 keV CT image, 1 feature from 80 keV CT image, 1 feature from 120 keV CT image, and 1 feature from 140 keV CT image) (Figures 3, 4).

In summary, five feature sets were formed, including the shape spectral CT set (n = 4), material-separation spectral CT set (n = 7), plain spectral CT set (n = 7), enhanced spectral CT set (n = 9), and combined feature set (n = 27) (Figure 4, Table 2).

On the cross-validation, the AUC of the shape model (AUC = 0.690, 95%CI: 0.622-0.753) was significantly smaller than that of the other four models (all P < 0.05). The combined model (AUC = 0.874, 95%CI: 0.821-0.916) outperformed the material-separation model (AUC = 0.828, 95%CI: 0.769-0.877, p = 0.005), the plain spectral CT model (AUC = 0.820 95%CI: 0.760-0.870, P = 0.005) and the enhanced spectral CT model (AUC = 0.838, 95%CI: 0.780-0.886, P = 0.005). There was no significant difference in performance between the material-separation model and the plain spectral CT model (P = 0.567), and between the material-separation model and the enhanced spectral CT model (P = 0.637), and the plain spectral CT model and the enhanced spectral CT model (P = 0.465) (Figure 5; Tables 3, 4). Within the risk threshold range of 0-0.8, the net benefit remained positive, demonstrating clinical utility. The combined model showed superior net benefit compared to other models across this threshold spectrum (Figure 6).