Diagnostic value of machine learning-based computed tomography texture analysis for differentiating multiple myeloma from osteolytic metastatic bone lesions in the peripheral skeleton.

OBJECTIVES

To report the diagnostic performance of machine learning-based CT texture analysis for differentiating multiple myeloma from osteolytic metastatic bone lesions in the peripheral skeleton.

METHODS

We retrospectively evaluated 172 patients with multiple myeloma (n = 70) and osteolytic metastatic bone lesions (n = 102) in the peripheral skeleton. Two radiologists individually used two-dimensional manual segmentation to extract texture features from non-contrast CT. In total, 762 radiomic features were extracted. Dimension reduction was performed in three stages: inter-observer agreement analysis, collinearity analysis, and feature selection. Data were randomly divided into training (n = 120) and test (n = 52) groups. Eight machine learning algorithms were used for model development. The primary performance metrics were the area under the receiver operating characteristic curve and accuracy.

RESULTS

In total, 476 of the 762 texture features demonstrated excellent interobserver agreement. The number of features was reduced to 22 after excluding those with strong collinearity. Of these features, six were included in the machine learning algorithms using the wrapper-based classifier-specific technique. When all eight machine learning algorithms were considered for differentiating multiple myeloma from osteolytic metastatic bone lesions in the peripheral skeleton, the area under the receiver operating characteristic curve and accuracy were 0.776-0.932 and 78.8-92.3%, respectively. The k-nearest neighbors model performed the best, with the area under the receiver operating characteristic curve and accuracy values of 0.902 and 92.3%, respectively.

CONCLUSION

Machine learning-based CT texture analysis is a promising method for discriminating multiple myeloma from osteolytic metastatic bone lesions.