Applying Machine Learning Analysis Based on Proximal Femur of Abdominal Computed Tomography to Screen for Abnormal Bone Mass in Femur

Abstract

To evaluate the performance of machine learninganalysis based on proximal femur of abdominal computed tomography (CT) scans in screening for abnormal bone mass in femur. 222 patients aged 50 years or older who underwent abdominal CT and dual-energy X-ray absorptiometryscans within 14 dayswere retrospectively enrolled. The patients were randomly assigned to a training cohort (n=155) and a testing cohort (n=67) in a ratio of 7:3. A total of 2288 candidate radiomic features were extracted from the volume region of interest- the left proximal femur of the abdominal CT scans. The most valuable radiomic features were selected using minimum-Redundancy Maximum-Relevancyand the least absolute shrinkage and selection operatorto construct the radiomics model. The predictive performance was assessed with receiver operating characteristiccurve. 13 features were chosen to establish the radiomics model. The radiomics model using logistic regressiondisplayed excellent prediction performance in distinguishing normal bone mass and abnormal bone mass, with the area under the curve (AUC), accuracy, sensitivity and specificity of 0.917 (95% CI, 0.867-0.967), 0.826, 0.935 and 0.780 in the training cohort. The testing cohort indicated a better performance with AUC, accuracy, sensitivity and specificity of 0.963 (95% CI, 0.919-0.999), 0.851, 0.923 and 0.889. The radiomics model based on proximal femur of abdominal CT scans had a high predictive performance to identify abnormal bone mass in femur, which can be used as a tool for opportunistic osteoporosis screening.