Automated detection of chronic kidney disease using image fusion and graph embedding techniques with ultrasound images

https://doi.org/10.1016/j.bspc.2021.102733Get rights and content

Highlights

  • Developed an efficient computer-aided diagnosis model to predict chronic kidney disease using ultrasound images.

  • Four-chamber heart Ultrasound images are employed to predict CKD stages.

  • Image fusion and graph embedding techniques are utilized.

  • The proposed method achieved an accuracy of 100 %, and 99.09 % for two-class and multi-class categorization respectively.

Abstract

Chronic Kidney disease (CKD) is a progressive disease affecting more than twenty million individuals in the United States. Disease progression is often characterized by complications such as cardiovascular diseases, anemia, hyperlipidemia and metabolic bone diseases etc., Based on estimated GFR values, the disease is categorized in 5 stages which significantly influence patient outcome. Cardiovascular ultrasound (US) (echocardiography) imagery demonstrate significant hemodynamic alterations that are secondary to CKD in the form of volume/ pressure overload. As the CKD pathology directly impacts cardiovascular disease, the US imaging shows structural and hemodynamic adaptation. Hence, the development of a computer-aided diagnosis (CAD) model to predict CKD would be desirable, and can potentially improve treatment. Several prior studies have utilized kidney features for quantitative analysis. In this paper, acquisition of the four-chamber heart US image is employed to predict CKD stage. The method combines image and feature fusion techniques under a graph embedding framework to characterize heart chamber properties. Moreover, a support vector machine is incorporated to classify heart US images. The proposed method achieved 100 % accuracy for a two-class system, and 99.09 % accuracy for a multi-class categorization scenario. Hence, our proposed CAD tool is deployable in both clinic and hospital settings for computer-aided screening of CKD.

Introduction

Chronic kidney disease (CKD) is said to be present when there is persistent kidney damage manifest as abnormal albumin excretion and/or reduced kidney function for more than 3 months. Kidney function can be quantified either by direct measurement or estimation of the glomerular filtration rate (GFR) [1,2]. CKD is an important public health problem that increases the mortality risk in affected patients [[3], [4], [5]]. Kidney disease has a major influence on global health, either as a direct cause of global morbidity and mortality, or as an associated risk factor for cardiovascular disease [3]. With increasing life expectancy and prevalence of lifestyle diseases, the mortality rate from CKD increased 41.5 % in the interval spanning from 1990 to 2017 [3]. To facilitate assessment of CKD severity, the National Kidney Foundation has developed criteria to stratify CKD patients into prognostic categories with ascending severity of CKD, as well as risks of further kidney disease progression: stage 1, kidney damage with normal or increased GFR (GFR ≥ 90 mL/min per 1.73 m2); stage 2, kidney damage with mild decrease in GFR (GFR of 60–89 mL/min per 1.73 m2); stage 3, moderate decrease in GFR (GFR of 30–59 mL/min per 1.73 m2); stage 4, severe decrease in GFR (GFR of 15–29 mL/min per 1.73 m2); and stage 5: kidney failure (GFR < 15 mL/min per 1.73 m2 or dialysis) [6].

Patients with CKD stages 1 and 2 are often asymptomatic and exhibit mild derangements in kidney function without clinically significant water, electrolyte, or endocrine/metabolic imbalances on serum assays. Hence, there is an imperative to develop safe, fast and cost-effective methods to screen for and to diagnose CKD accurately, so that early preventive measures can be initiated to retard the deterioration of kidney function. In patients at all stages of CKD, widely available kidney US can be performed at relatively low cost to assess for structural changes associated with CKD, such as reduced kidney size or change in texture of the kidney parenchyma. The latter requires qualitative visual analyses of US imagery by expert radiologists, which is time-consuming and subject to inter-observer variability [7]. Echocardiographic imaging may reveal chamber dilatation, systolic/diastolic dysfunction of both atria and ventricles, hypertrophy of left ventricular wall, early degenerative valvular heart disease resulting in valve stenosis, and pulmonary hypertension with deteriorated right ventricular function. Patients with higher stages of renal failure differ from the healthy population by cyclic changes of hydration. In the normal condition, the kidney regulates the body fluid volume and controls water homeostasis whereas there will be cyclic changes in hydration in CKD. Profound cardiac changes are more evident in higher stages of CKD [47]. Echocardiography, being a feasible method for screening and assessment of cardiac geometry and function, exhibits subjectivity. The cardiac MRI is thought to be the gold standard in this field to the present time [[48], [49], [50]].

Several investigators have proposed computer-aided diagnosis (CAD) systems for CKD detection and for severity assessment using various combinations of clinical, serological, and kidney US imaging parameters as inputs. Building on 10-year data of new CKD patients, Norouzi et al. [8] developed a fuzzy expert system that could predict future GFR at 6, 12, and 18 months with greater than 95 % accuracy by modeling weight, diastolic blood pressure, diabetes mellitus status, and estimated GFR level. Samir et al. [9] investigated the use of an US shear wave elastography technique (SWE) among 25 subjects with stage 3 or 4 CKD versus 20 healthy controls, and observed that Young’s modulus assessed by SWE was discriminative for the stiffer renal tissue in CKD subjects versus normals. This finding was corroborated by Leong et al. [10] in a larger study of 106 CKD subjects and 203 controls, in which use of a cut-off value of Young’s modulus greater than 4.31 kPa outperformed renal length and cortical thickness measurements for separating CKD versus non-diseased kidney tissue. In another study involving 25 CKD subjects and 10 controls, multiparametric quantitative US imaging comprising anatomical, Doppler, shear wave, and image pixel-intensity analytic elements demonstrated correlation of some of these elements with estimated GFR-assessed mild CKD, moderate to severe CKD and normal classes, as well as histopathological findings in the CKD subjects who all underwent kidney biopsy [11]. Almansour et al. [12] studied machine learning approaches for CKD detection on a 400-subject cohort with 24 attributes (i.e., nominal and numerical) and found that an artificial neural network (ANN) had the best classification accuracy at 99.75 %. Hao et al. [7] proposed a novel deep learning method called texture branch network that automatically selected and fused kidney US textural features, and achieved a high classification accuracy of 96.01 %. Acharya et al. [13] studied 405 kidney US images and employed elongated quinary patterns and bispectrum to detect CKD with 99.75 % accuracy. Kolachalama et al. [14] implemented a deep convolutional neural network (CNN) for classification of renal biopsy images that outperformed pathologist-estimated fibrosis scores for association with CKD stage, serum creatinine, and proteinuria, as well as prediction of survival. Recently, Ma et al. [15] used adaptive, histogram, and Haralick features to characterize normal versus abnormal kidney US images, and ANN to classify the features.

CKD is associated with a higher incidence of cardiovascular disease, hyperlipidemia, anemia, and metabolic bone disease, which in turn lead to serious downstream complications [16]. In dialysis-dependent stage 5 CKD patients, the risk of cardiovascular mortality is increased by 10–20 times compared with control subjects without CKD [17]. Indeed, cardiovascular pathology can be present in CKD patients even before dialysis dependence. Various cardiac structural and functional changes have been documented among CKD patients, including increased left ventricular mass, diastolic and systolic left ventricular dysfunction; and in the clinical setting, heart failure is associated with the higher stages of CKD [18]. Continual cardiac assessment of CKD patients throughout the course of disease is therefore necessary for timely identification of cardiovascular complications and preemptive optimization of medical treatment, to reduce morbidity and mortality. In this study, we propose to investigate cardiac structural changes among patients with stage 3, 4, and 5 CKD versus healthy controls. Specifically, a CAD model will be applied to apical four-chamber echocardiographic US images acquired from CKD and healthy controls/normal individual subjects.

The novelty and key contributions of this paper are as follows:

  • We differentiated cardiac imagery of stage 3–5 with healthy controls/normal individuals.

  • Image and feature level fusion approaches were developed and implemented to characterize textural information of heart US imagery at various stages of CKD.

  • The local neighborhood is preserved with the help of supervised graph embedding by integrating class label information.

  • A parsimonious approach using only the cardiac apical four-chamber US view is employed to streamline the analysis. To our knowledge, this is a novel approach to correlate cardiac pathology with CKD stage. The developed algorithm is useful for extracting salient information from heart US images, and can potentially be generalized to other medical imaging domains and analytic applications.

The remainder of the paper is arranged in the following order: Section 2 describes the data acquisition and delineates the proposed method to estimate CKD severity. The experimental results and comparative review of various extant methods are presented in Section 3. A discussion and conclusion are provided in Sections 4 and 5, respectively.

Section snippets

Data description

The study population comprised of 120 patients with CKD stage 3–5 (46 stage 3, 41 stage 4 and 33 stage 5) diagnosed based on an estimated GFR value ≤59 mL/min per 1.73 m2 who consulted at the Cardiology outpatient department, Kasturba Hospital, Manipal, Karnataka, India, along with 100 age-matched controls. CKD stage 1 and 2 patients were excluded, as they are less likely to harbor cardiac manifestations secondary to CKD. The study protocol was approved by the institutional ethics committee,

Experimental setup

The complete algorithm is implemented under MATLAB and executed using a personal computer with Intel Core-i5 6200U and 4 GB RAM memory. To achieve optimal performance, the required parameters are set to predefined empirical values as shown in Table 1.

We divided the dataset into 10 partitions, where 1 partition was for testing and the remaining 9 partitions were for training. Every partition was tested by assembling the remaining 9 partitions as the training data, and finally the average

Discussion

This paper aims to develop an automated system to categorize different stages of CKD. The proposed method uses a robust feature extraction technique, which is the combination of image and feature level fusion techniques. The clinical features of the heart are extracted using steerable filters and fusion techniques efficiently, as features are highlighted every 10 degrees. This helps to enhance the periphery of the heart. The variation of textural pattern is captured using only four entropy

Conclusion

In this paper, estimation of CKD is proposed using an efficient feature extraction technique under a graph embedding framework. The graph embedding is able to differentiate CKD staging by preserving local inherent characteristics. Our experimental results show that the proposed method performs well in classifying CKD staging. The combination of fusion and graph embedding overcome the limitation of the unbalanced class frequency problem and shows excellent classification accuracy. The system is

Funding

NA.

Availability of data and material

NA.

Code availability

Custom code.

CRediT authorship contribution statement

Anjan Gudigar: Conceptualization, Methodology, Software, Data curation, Writing - original draft, Visualization, Investigation. U Raghavendra: Conceptualization, Methodology, Software, Data curation, Writing - original draft, Visualization, Investigation. Jyothi Samanth: Conceptualization, Methodology, Software, Visualization, Investigation. Mokshagna Rohit Gangavarapu: Software, Data curation. Abhilash Kudva: Software, Data curation. Ganesh Paramasivam: Visualization, Investigation.

Declaration of Competing Interest

The authors report no declarations of interest.

References (51)

  • D.F. Specht

    Probabilistic neural networks

    Neural Netw.

    (1990)
  • Aapo Hyvärinen et al.

    Independent component analysis: algorithms and applications

    Neural Netw.

    (2000)
  • Svante Wold et al.

    Principal component analysis

    Chemom. Intell. Lab. Syst.

    (1987)
  • Y.W. Liu et al.

    The role of echocardiographic study in patients with chronic kidney disease

    J. Formos. Med. Assoc.

    (2015)
  • J. Coresh et al.

    Prevalence of chronic kidney disease and decreased kidney function in the adult US population: third national health and nutrition examination survey

    Am. J. Kidney Dis.

    (2003)
  • KDOQI clinical practice guidelines and clinical practice recommendations for anemia in chronic kidney disease

    Am. J. Kidney Dis.

    (2006)
  • Boris Bikbov et al.

    Global, regional, and national burden of chronic kidney disease, 1990–2017: a systematic analysis for the Global Burden of Disease Study 2017

    Lancet

    (2020)
  • M. Ruiz-Ortega et al.

    Targeting the progression of chronic kidney disease

    Nat. Rev. Nephrol.

    (2020)
  • Thomas Ebert et al.

    Inflammation and premature ageing in chronic kidney disease

    Toxins

    (2020)
  • P. Hao et al.

    Texture branch network for chronic kidney disease screening based on ultrasound images

    Front. Inf. Technol. Electron. Eng.

    (2019)
  • Jamshid Norouzi et al.

    Predicting renal failure progression in chronic kidney disease using integrated intelligent fuzzy expert system

    Comput. Math. Methods Med.

    (2016)
  • Anthony E. Samir et al.

    Shear wave elastography in chronic kidney disease: a pilot experience in native kidneys

    BMC Nephrol.

    (2015)
  • S.S. Leong et al.

    Shear wave elastography in the evaluation of renal parenchymal stiffness in patients with chronic kidney disease

    Br. J. Radiol.

    (2018)
  • Jing Gao et al.

    Multiparametric quantitative ultrasound imaging in assessment of chronic kidney disease

    J. Ultrasound Med.

    (2017)
  • U.R. Acharya et al.

    Automated detection of chronic kidney disease using higher-order features and elongated quinary patterns from B-mode ultrasound images

    Neural Comput. Appl.

    (2020)
  • Cited by (7)

    View all citing articles on Scopus
    View full text