A Novel Approach Based on Integrating Radiomics, Bone Morphometry and Hounsfield Unit-Derived From Routine Chest Ct for Bone Mineral Density Assessment Publisher

Summary: Study shows combining CT scan features with machine learning accurately detects bone density issues, aiding osteoporosis diagnosis. #BoneHealth #Radiology

Mohammadisadr M¹ ; Cheki M² ; Moslehi M¹ ; Zarasvandnia M² ; Salamat MR¹
Source: Academic Radiology Published:2024

Abstract

Rationale and Objectives: The purpose of this study is the feasibility of using radiomics features, bone morphometry features (BM), and Hounsfield unit (HU) values obtained from routine chest computed tomography (CT) for assessing bone mineral density (BMD) status. Materials and Methods: This retrospective study analyzed 120 patients who underwent routine chest CT and dual-energy X-ray absorptiometry examinations within a month. Whole thoracic vertebral bodies from routine chest CT images were segmented using the GrowCut semi-automatic segmentation method, and radiomics features, BM features, and HU values were extracted. To assess the intra- and inter-observer variability of segmentation, the Dice similarity coefficient (DSC) was utilized. Feature selection was carried out using the intra-class correlation coefficient and the Boruta algorithm. Six machine learning classification models were employed for classification in a three-class manner. The models’ performance was evaluated using the area under the receiver operator characteristics curve (AUC). Other evaluation parameters of the models were calculated, including overall accuracy, precision, and sensitivity. Results: The DSC values showed high similarity by achieving 0.907 ± 0.034 and 0.887 ± 0.048 for intra- and inter-observer segmentation agreement, respectively. After a two-stepwise feature selection, 21 radiomics features were selected. Different combinations of these radiomics features with five BM features and HU values were applied to six classification models for evaluating BMD. The multilayer perceptron (MLP) model based on integration of radiomics features and BM features in a three-class classification approach achieved higher performance compared to other models with an AUC of 0.981 (95% confidence interval (CI): 0.937–0.997) in normal BMD class, an AUC of 0.896 (95% CI: 0.826–0.944) in osteopenia class, and an AUC of 0.927 (95% CI: 0.866–0.967) in osteoporosis class. Conclusion: Using the MLP classification model based on a combination of radiomics features and BM features in a three-class classification approach can effectively distinguish different BMD conditions. © 2024 The Association of University Radiologists