Power transition signatures of vibroarthrographic spectrograms for diagnosing knee joint pathologies

Vibroarthrographic (VAG) signals are extensively used for the diagnosis of knee joint pathologies. Since wear and tear of the cartilage affects the packing density of knee joints, pathological VAG signals are less non-stationary than healthy VAG signals. State-of-the-art feature descriptors extracted from the time-frequency (TF) representations of VAG signals do not have sufficient discrimination power to distinguish normal and abnormal VAG signals. To resolve this issue, we propose two novel features, mean of temporal dispersion of power at frequencies $\left(\text{MTDPF}\right)$ and mean of successive temporal difference in power across frequencies $\left(\text{MSTDPF}\right)$ computed from the spectrogram to quantify the non-stationarity of VAG signals. While $\text{MTDPF}$ is the standard deviation of the spectral power across time, averaged among the frequency bins, $\text{MSTDPF}$ is the average of absolute differences in spectral power across all time windows and frequency bins. $\text{MTDPF}$ and $\text{MSTDPF}$ are observed to be significantly different between normal and abnormal VAG signals with p-values of 0.00032 and 0.00070 respectively. $\text{MTDPF}$ and $\text{MSTDPF}$ are observed to be more determinant compared to the Gray Level Co-occurrence Matrix (GLCM) features and local binary patterns (LBP) features of spectrogram itself and a broad spectrum of non-stationary features extracted from other TF representations. So, these two features $\text{MTDPF}$ and $\text{MSTDPF}$ can be employed in the automated systems used for the interpretation of VAG signals.