Abstract

Osteoporosis is a skeletal disorder characterized by bone loss and increased risk of fragility fractures (FF). It's clinically defined based on the current gold standard, Dual-energy X-ray Absorptiometry (DXA). Axial Transmission (AT), an ultrasonic device, has been proposed as a DXA alternative. Recently, authors have proposed using raw radiofrequency signals (RS) collected with an AT device and a multi-channel convolutional neural network (MCNN) to classify patients with or without FF. This study aims to improve the classification using a different AT device RS and an optimized MCNN. A previous clinical study database was used to achieve this objective, with 14,937 raw signals from 195 patients, 91 with FF and 104 without FF. Patients were split into train (70%) and validation (30%). The MCNN architecture contains 5 blocks: convolution, max pooling, drop out, and batch normalization layers. A flattened vector is completed using the following clinical data: age, BMI, and cortisone intake. The results using 5-fold Stratified cross-validation showed an average accuracy of 0.76 and loss of 0.72, which was above the accuracy obtained with DXA (acc = 0.65) and the same patients.