Evaluating bone density with ultrasonic backscatter: Leveraging time-frequency analyses and convolutional neural...

Osteoporosis is a disease characterized by decreased bone density and increased fracture risk. This study proposes a convolutional neural network (CNN) method for detecting changes in bone density using spectrograms or scalograms generated from ultrasonic backscatter signals of 55 human cancellous bone specimens from the proximal femur. After applying a simple grid search to optimize three CNN hyperparameters (learning rate, batch size, and number of epochs), the CNN model was able to accurately determine bone density from the ultrasound data. Linear regression analysis revealed strong correlations between the densities predicted by the CNN model and the actual densities of the specimens (R² = 0.98 and R² = 0.94 for the spectrogram and scalogram inputs, respectively) with a slope close to one in both cases. These results suggest that a CNN model based on spectrograms and scalograms derived from ultrasonic backscatter signals from cancellous bone may provide a sensitive method for detecting osteoporotic changes in bone, outperforming conventional backscatter methods. Spectrograms are found to offer slightly better performance than scalograms as the CNN model input. The performance of the CNN model is robust, exhibiting little dependence on fine-tuning of the hyperparameters and the size of the training set.