1.1 Patients
This retrospective study was approved by the Institutional Review Committee and informed consent was waived. Patients who underwent lumbar CT and DXA from January 2020 to August 2021 were searched in our hospital database. Inclusion criteria were as follows :(a) patients underwent both lumbar CT and DXA; (b) The average interval between all examinations < 14 days.
Exclusion criteria :(a) patients with various abnormal bone metabolic diseases; (b) patients with vertebral bone tumor and metastasis; (c) patients with hematological disorders; (d) Poor image quality; (e) Large bone islands of vertebral body; (f) Patients with lumbar fractures, deformation, and vertebra surgery. Finally, 170 patients met the criteria for inclusion in the study and were randomly assigned to the training and validation groups in a 7:3 ratio. Figure 1 illustrates the different steps of the methodology followed in this work.
1.2 DXA examination
DXA of the lumbar spine (L1-L4) and left proximal femur (including femoral neck, greater trochanter, and Ward's triangle) for bone mineral density(BMD) assessment was performed using a single BMD scanner (OsteoSys,Global BMD,PRIMUS-Application).Patients were classified into three categories according to the criteria: osteoporosis ( T-score ≤ -2.5 ), osteopenia (-2.5 < T-score < -1.0) and normal (T-score ≥ -1.0)[9].
1.3 CT Examination
All study participants underwent CT of the spine on a helical 256 channel CT scanner(Revolution;GE Healthcare), Scanning conditions: tube voltage 120kVp, tube current of 355 mAs, layer thickness 5 mm, layer spacing of 5 mm, (window width/level, 2000/300).The patients were scanned in supine position from the L1 to L4 vertebral body,Two - dimensional reconstruction images were obtained on coronal and sagittal planes.
1.4 ROI segmentation
After CT scanning, L1-L4 axial bone window images were reconstructed with a layer thickness and layer spacing of 1.25mm. The region of interest (ROI) was manually drawn on the axial images by a radiologist with more than 10 years of CT diagnostic experience using the ITK-SNAP software. The ROI was acquired in the L1-L4 vertebral bodies (below the upper endplate, in the middle of the vertebral body, and above the lower endplate). The boundary was set along the inner edge of the vertebral cortex. Meanwhile, bone cortex, and vertebral venous plexus and osteoproliferation were avoided, while the vertebral bodies were included as much as possible (Fig. 2).Images containing ROI information are saved in NIfTI format for subsequent feature analysis.The image segmentation was supervised by another experienced radiologist,and the difference in opinion was resolved through consultation.
1.5 Radiomic features extraction
Radiomic features were extracted using AK software (AnalysisKit, version 3.2.0, GE Healthcare, China) and pyradiomicshttps(version 3.0.1,https://pyradiomics.readthedocs.io/en/latest/).Before feature extraction, the pictures were preprocessed in three steps: resampling the voxel size to 1×1×1 mm3, discretizing the gray values with a bin width of 25, and normalizing the gray value. Finally, 7 classes of features were obtained. First Order, shape, gray level size zone matrix (GLSZM), gray level run length matrix (GLRLM),gray level dependence matrix (GLDM), Gray level co-occurrence matrix (GLCM), neighbouring gray tone difference matrix (NGTDM),based on the original image, laplacian of gaussian filtered images with sigma 2, 3.
1.6 Radiomic Model Development
We used two feature selection methods, the Max-Relevance and Min-Redundancy (mRMR) and the least absolute shrinkage and selection operator (LASSO) to select the features.Firstly, mRMR was used to remove redundant and unnecessary features,30 features were kept. Then, LASSO was used to choose the best subset of features from which to build the final model. The LASSO initially used 10-fold cross-validation to tune the regular parameter λ, and the optimized λ corresponded to the least amount of model bias. Then, using the optimized λ, the number of features was obtained, the best predictive subset of features was selected, and the associated coefficients were assessed. The radiomics signature (Radscore) was computed by adding the weighted total of the relevant characteristics.Each subject's Rad-score was determined by applying the following equation:
\(Rad-score=\sum _{i=0}^{n}ci\times\) xi་b
where b is the intercept, Ci is the associated feature coefficient, and Xi is the ith selected feature.
1.7 Statistical Analysis
The R studio (version 4.0.2,www.r-project.org) and IBM SPSS Statistics 25.0 were utilized for all statistical analyses. One-way ANOVA test was used to analyze the differences in clinical continuous variables (including age, weight, and height).The machine learning classification models were performed using R software. The area under ROC curve (AUC), accuracy, specificity, and sensitivity were used to evaluate the effectiveness of identifying abnormal bone mass in each group.V alues of p ≤ 0.05 were considered as statistically significant .