Utilizing radiomics techniques to isolate a single vertebral body from chest CT for opportunistic osteoporosis screening

Abstract

Purpose: Opportunistic osteoporosis screening, conducted during routine medical examinations such as chest computed tomography (CT), presents a potential solution for early detection. This study aims to investigate the feasibility of utilizing radiomics technology based on chest CT images to screen for opportunistic osteoporosis.

Methods: This Study is a Multicenter Retrospective Investigation. Relevant clinical data, including demographics and DXA results, would be collected for each participant. The radiomics analysis in this study focuses on the extraction of features from the 11th or 12th thoracic vertebral bodies from chest CT images. SVM machine learning models would be trained using these radiomic features, with DXA results as the ground truth for osteoporosis classification.

Results: In the training group, Clinical models had an accuracy of 0.684 and an AUC of 0.744, Radiomics models had an accuracy of 0.828 and an AUC of 0.896, Nomogram models had an accuracy of 0.839 and an AUC of 0.901. In the internal validation group, Clinical models had an accuracy of 0.769 and an AUC of 0.829, Radiomics models had an accuracy of 0.832 and an AUC of 0.892, Nomogram models had an accuracy of 0.839 and an AUC of 0.918. In the external validation group, Clinical models had an accuracy of 0.715 and an AUC of 0.741, Radiomics models had an accuracy of 0.777 and an AUC of 0.796, Nomogram models had an accuracy of 0.785 and an AUC of 0.807. In all three datasets, the Nomogram model exhibited a statistically significant difference in screening effectiveness compared to the clinical models.

Conclusion: Our research demonstrates that by leveraging radiomics features extracted from a single thoracic spine using chest CT, and incorporating these features with patient basic information, opportunistic screening for osteoporosis can be achieved.

Keywords: Chest CT; Osteoporosis; Radiomics; Thoracic vertebrae.

Fig. 1

The process of patient inclusion

Fig. 2

The process of outlining ROI

Fig. 3

The process of feature extraction and selection. A. Visualization results of extracted radiomics features. B. Cluster analysis of extracted radiomics features

Fig. 4

The process of feature extraction and selection. A. Classification of radiomics features. B. C. LASSO regression. D. The final selected features

Fig. 5

A.The ROC curves of the three models in the training group. B. The DCA curves of the three models in the training group. C. The calibration curves of the three models in the training group

Fig. 6

A.The ROC curves of the three models in the training group. B. The DCA curves of the three models in the training group. C. The calibration curves of the three models in the training group

Fig. 7

A.The ROC curves of the three models in the training group. B. The DCA curves of the three models in the training group. C. The calibration curves of the three models in the training group. D. A clinical column chart that can be used for clinical prediction