An exploratory study of explainable deep learning for predicting bone mineral density using clavicle features on chest radiographs: A multi‐task approach with regression and segmentation

Abstract

Abstract Purpose Although bone mineral density (BMD) measurement using dual‐energy x‐ray absorptiometry (DXA) is the most common method of diagnosing osteoporosis, it is not widely used to screen patients. In this exploratory study, we developed a multi‐task learning model that predicts BMD from chest radiographs using clavicular features and supports network explainability. Methods The proposed multi‐task learning model integrates segmentation and regression tasks by incorporating a regression branch into the U‐Net architecture in an end‐to‐end manner. A total of 1600 patients who underwent chest radiography and DXA of the lumbar vertebrae were included in this study. We compared the BMD predictive performance of the mean absolute error (MAE) and Pearson correlation coefficient ( R value) between the proposed multi‐task learning model and the single‐task learning model, which was defined as the comparison model that only performed BMD prediction. Additionally, model performance for classifying osteoporosis, osteopenia, and normal bone status was evaluated via reclassification analysis based on the World Health Organization (WHO) criteria. Confusion matrices were generated, and classwise and macro‐averaged performance metrics were calculated. To confirm the rationale for the BMD predictions, we evaluated heat maps generated using the gradient‐weighted class activation mapping technique to determine whether the highlighted regions overlapped with the clavicle. Results The multi‐task learning model demonstrated superior predictive performance (MAE of 0.092 g/cm 2 and R value of 0.769) compared with the single‐task learning model (MAE of 0.101 g/cm 2 and R value of 0.724), a statistically significant ( p < 0.001) difference in MAE. Bland–Altman analysis showed that the multi‐task learning model had good agreement with narrower limits of agreement, although a bias was present (bias: −0.013 g/cm 2 ; limits of agreement: −0.248 to 0.223 g/cm 2 ). In contrast, the single‐task model showed slightly wider agreement limits (bias: −0.003 g/cm 2 ; limits of agreement: −0.257 to 0.252 g/cm 2 ). In the reclassification analysis based on the WHO criteria, the multi‐task learning model resulted in fewer misclassifications than the single‐task learning model. The macro‐averaged sensitivity, specificity, precision, and F1 score were 0.647, 0.826, 0.680, and 0.659, respectively, for the multi‐task model, compared with 0.597, 0.809, 0.660, and 0.616, respectively, for the single‐task model. The heat maps in the multi‐task learning model highlighted different regions compared with the single‐task model, the clavicular area. Conclusion The proposed multi‐task learning model demonstrated the predictive rationale by focusing on the clavicle in chest radiographs, which is clinically relevant to BMD, and showed improved performance compared with the single‐task model.

Ähnliche Arbeiten