Time-Frequency Analysis of Vibroarthrographic (VAG) Signals and Classification based on Deep Learning

Abstract

The vibration and sound pattern generated by knee joints during their movement reflects their condition. The pattern may vary due to any injury, damage, or degeneration. For further clinical treatment, it is important to diagnose early any knee joint disorders. Several studies based on time-frequency parameters and statistical modelling techniques have been proposed for analyzing and classifying VAG signals, but no prominent work has been done in this regard by using Deep learning techniques. In the analyses of medical images, deep learning algorithms are proved to be very successful due to which Convolutional Neural Networks (CNNs) have gained much attention. In the research work that is being presented here, a new Deep Learning based method combined with supervised Machine learning pattern classification algorithms is presented for analyzing and classifying VAG signals. The time-frequency analysis of VAG signals is performed via Short-time Fourier transform (STFT) and that time-frequency distribution is considered as a 2D image of time-frequency. We have also proposed to use pretrained Convolutional Neural Network (CNN) models for feature extraction. Further, Principal Component Analysis (PCA) algorithm is employed for the selection of significant features, for reducing the dimensions of feature vector and to increase interpretability. Support vector machine (SVM) and Random Forest both were used for classification of healthy and pathological subjects. Statistical results obtained concludes that the proposed method can classify VAG signals with a classification accuracy of 92.13%, specificity of 92.16%, sensitivity of 92.11%, and a Matthews Correlation Coefficient (MCC) score of 0.84. The proposed method can accurately screen around 98% of subjects based on area under the ROC curv