Analysis of knee-joint vibroarthographic signals using bandwidth-duration localized three-channel filter bank

Highlights

• Novel optimal bandwidth-duration localized three band wavelet filter bank based features extracted from vibroarthographic signals.

• Attained very good classification performance in the identification of anomalies related to knee joints.

• An cost-effective and non-invasive computer-aided diagnosis system.

• Employed only one feature-set of log energy.

Abstract

A vibrarthographic (VAG) signal is the recorded vibration signal produced by human knee-joints during movements of the leg. These signals are useful in quantifying lubrication and roughness of articular cartilage layers of the knee-joint. The early detection and treatment of an abnormal knee-joint will help in providing a better quality of life for the suffering patients. The VAG signals are nonlinear and non-stationary. Hence, in the proposed study, we have employed an optimal bandwidth-duration localized three-band (OBDLTB) orthogonal wavelet filter banks (OWFB) to automatically identify the health of knee-joints using VAG signals. We have used the database created at the University of Calgary containing a total of 89 VAG signals (51 normal and 38 pathological). The log-energy (LOEN) features are obtained from various sub-bands using OBDLTB-OWFB. The extracted LOEN features are separated using supervised machine learning algorithms into normal and abnormal VAG signals. Our automated system obtained an average classification accuracy of 89.89% and F1-score of 87.67% with 10-fold cross-validation. The performance of the model indicates that the proposed technique is able to identify the abnormal knee-joint reliably and thus this technique can aid the orthopedic surgeons in the early diagnosis of knee-joint abnormalities.

Introduction

The knee-joints of human beings support the entire weight of the human body are intricate and strong. The knee-joint is made up of three parts: femur, tibia, and patella. Due to aging the bones degenerates and causes discomfort and pain, leaving one to undergo knee replacement without any alternative. Osteoarthritis is one such common disease which is associated with degradation of articular cartilage. The degrading joint disease is most prevalent in aged people. It affects the joints like knees, hips, fingers, and lower spine regions which have been stressed for a long time. The osteoarthritis in developed countries is one among ten most disabling diseases [1]. Statistics show that 9.6% of males and 18.0% of females across the world with more than 60 years of age are suffering from symptomatic osteoarthritis [1].

The operative procedure of knee replacement includes the surgical treatment of damaged and worn out surfaces of the knee-joint [2]. Knee replacement is often done for various knee diseases like osteoarthritis, psoriatic arthritis, and rheumatoid arthritis. Conventionally, this operation is performed by replacing the lower part of the femur and bone at the upper part of the tibia. Partial and complete knee replacement can be surgically performed through this technique. The metal and plastic made elements are designed to replace the worn out joint surfaces with proper design parameters set for the smooth movement of the knee. For the pain-free movement of the leg, it is required to replace the affected surfaces of the knee-joint with these specially molded metal and plastic components, which is expensive and therefore the early detection of the disorder becomes very important. Various traditional imaging procedures like computer tomography, X-Ray and magnetic resonance imaging (MRI) do not provide accurate results in the early stages of the disease as they are unable to detect the minute changes. The popularly used gold standard for the examination of the state of knee-joint is arthroscopy in which the cartilage surface is probed by inserting a fiber-optic cable along with a tiny device that consists of a lens and illuminating systems for magnifying and illuminating interior condition of cartilage to notice any possible inflammation, chondromalacia ,tears of meniscus and anterior cruciate ligament tears. The limitation of arthroscopy is that we cannot employ it in the cases where patients have highly degenerated knee-state because of disorders such as osteoarthritis, meniscectomy and patellectomy. In some situations, as described in [3], even though the joint deteriorates, the arthroscopy test may not be able to identify it. At the same time, it should be noted that arthroscopy is a semi-invasive test. Hence it is also not practical to detect the knee-joint disorder.

An accurate noninvasive knee-joint anomaly (KJA) detection can be achieved through the vibroarthographic (VAG) signal. The VAG signal is the recorded acoustic-vibration signal generated from the knee-joints of human beings due to natural actions of the leg. The VAG signal is usually obtained as described in [4] where the person is positioned on a fixed table in a un-strained posture with his/her leg being freely suspended.

These sound signals are generally produced around mid-patella area and can be obtained with the help of an accelerometer while the person moves his/her leg from full extension state to complete bending state [5]. Hence a VAG signal has two parts, one part relates to the bending and the other half to the extension. Some of the salient properties of VAG signals are listed below [6]:

• They are nonlinear non-stationary because during the articulation of the joint, the quality of the joint contact-surfaces changes with the angular position of the leg.

• The amplitude and frequency characteristics of VAG signals are different for normal and abnormal classes [3].

• It is possible that a VAG signal obtained is a multicomponent signal because when the leg is moving, the friction between femoral condyle and layer over patella results in aggregate of multiple vibrations.

• An apriori estimation of signal to noise ratio of vibration signals is impossible as the random noise gets added to the signal due the movements during data acquisition.

Hence, we cannot use naive signal processing approaches to examine non-stationary VAG signals. The approaches like auto-regressive (AR) modeling also cannot precisely process the VAG signal. In AR modeling based analysis, non-stationary VAG signal is divided adaptively in stationary segments. Thus, in AR modeling, a non-stationary VAG signal is represented with stationary AR parameters assuming the segment of the signal is stationary, which may lead to an inaccurate or false analysis of the signal. The application of finite-duration segmented VAG signals may cause inaccurate estimation of the spectrum or power spectral density of the signal in low-frequency ranges. Hence, the segmentation-based AR methods are not attractive for mapping the clinical information acquired, using arthroscopic test, from the underlying segmented VAG signal. In order to overcome these constraints, advanced signal processing techniques which take into non-stationary behavior are highly desirable. Hence the wavelet-based non-stationary method is employed in this work.

Many researchers have made significant contributions using of VAG signals namely: processing [7], improve the quality of the signal by canceling muscle interference and developing parametric representation and classification [5], [8]. These vibration based signals are now being used to screen the patients with knee problems, before getting them to the arthroscopic examination. This need is proven, as most of the patients undergone arthroscopy were erroneously diagnosed free from any joint disorder [8]. This will help as a screening procedure for physicians and orthopedic specialists. To achieve this objective, the various feature-related parameters are investigated for the categorization of VAG signals into two classes (normal and abnormal). The VAG signals are analyzed in their entirety, without segmentation of parts of the signals, which helps in reducing the complexity in signal processing.

Several research works have been performed related to the discrimination of VAG signals. Umapathy and Krishnan [9] have obtained a classification accuracy of 79.8% with wavelet decomposition and involving local discriminant bases. Rangayyan and Wu [10] employed statistical and entropy-based features for the classification of normal and abnormal VAG signals. In another method [11], they have employed fractal dimension FD based features. Recently, Sharma et al. [12] have used features that are obtained from VAG signal by decomposing them via double-density dual-tree wavelet transform. The bandwidth-duration localized wavelets have been proven to be highly useful in analyzing non-stationary physiological signals [13], [14]. Further, the performance of three-band wavelet-based features has been found to be superior to the commonly used two-band wavelet-based features [15]. This leads us to explore the performance of bandwidth-duration optimized three-band filter banks (FBs) in the automated KJA detection system using VAG signals. In this work, we introduced a new approach for screening of VAG signals using optimal bandwidth-duration concentrated three-band (OBDLTB) orthogonal wavelet filter bank (OWFB). The log-energy (LOEN) features of sub-bands (SBs) of VAG signals are exploited as the distinguishing features. Firstly, VAG signals have been pre-processed to remove noise and artifacts, and then the preprocessed signals are decomposed using OBDLTB-OWFB into seven sub-bands. The LOENs are computed for each sub-band. We employed the t-test to rank the extracted features. Following this, the features are input to the least squared-support machine (LS-SVM) algorithm for the classification of normal and abnormal classes.

The objective of this study is to develop a simple, accurate and computationally less expensive automated system for knee-joint screening. For the automated detection of KJA, we have used an optimally duration-bandwidth concentrated three-band orthogonal wavelet filter bank (OWFB). Previously used AR modeling based methods [3], [5], [6] cannot be considered as the best candidate for the analysis of non-stationary signals including VAG signals because of their limitations as mentioned in [11]. Wavelet-based techniques and joint time-frequency localization based tools are more suitable for time-frequency analysis of VAG signals [9], [12], [16], [17]. We have used a novel wavelet-based technique for analyzing the VAG signals in the investigation. The features derived from three-band OWFB have not been used before despite the advantages of three-band OWFBs [18] over conventional two-band filter banks [19] in analyzing the non-stationary signals. In this work, two mother wavelets and a father wavelet (scale function) are generated by the iteration of the three-band filter bank. The filters are optimized for their joint bandwidth duration concentration as the joint concentration is highly desirable for the accurate analysis of non-stationary signals. Thus, novel OTFLTB-OWFB-based features are employed to discriminate two classes of VAG signals.

The rest of the paper has been organized as follows: Section 2 describes the database used in this study. In Section 3, we explain the methodology used for the proposed automated knee joint anomaly detection system. Section 4 presents the classification results and also discusses certain salient features of the KJA detection system designed by us. Section 5 concludes the presented work.