Elsevier

Measurement

Volume 153, 1 March 2020, 107426
Measurement

Analysis of quantum noise-reducing filters on chest X-ray images: A review

https://doi.org/10.1016/j.measurement.2019.107426Get rights and content

Highlights

  • Effects of Quantum noise on CXR images and performance of CAD systems.

  • Review and analyzed various benchmark filters for reducing noise in CXR images.

  • Investigated the tradeoff between de-noising and texture preserving performance.

  • Guided filter outperformed in terms of quantitative, and subjective measures.

  • The statistical evaluation proved the significance of the obtained results.

Abstract

Radiography is one of the important clinical adjuncts for preliminary disease investigation. The X-ray images are corrupted with inherent quantum noise affecting the performance of computer-aided diagnosis systems. This paper presents an extensive experimental review and impact of six benchmark filters for reducing noise and disease classification on chest X-ray images. The tradeoff between de-noising and texture preserving performance is investigated through classification performances using the state-of-the-art machine learning methods – Support Vector Machine and Artificial Neural Network. Moreover, the qualitative, subjective, and statistical evaluation is performed by using the image quality metrics, expert radiologist opinion, and statistical test, respectively. The experimental results confirm the significant improvement in classification performance using Guided filtered images. Furthermore, the results of qualitative measures and subjective analysis demonstrate that the guided filter and anisotropic diffusion filter both performed significantly better. Finally, a non-parametric statistical test is used to validate statistical significance of the obtained results.

Introduction

Automated medical image analysis is one of the challenging tasks due to inherent noise in medical imaging modalities, complex and overlapping anatomical structures, and vague appearance of disease manifestations. X-ray images are low-cost, globally accepted medical imaging modality used as a primary clinical adjunct for disease interpretation and accurate pathological decision making. Recent advancement in medical imaging technologies generates millimeter-scale resolution images to gain the insight of internal anatomical structures suitable for texture analysis used in computer-aided diagnosis (CAD) based image-guided therapy [1]. However, these images are predominantly corrupted from inherent Quantum noise, referred to as Photon noise, Poisson noise, or Shot noise [2]. The quantum noise creates random abrasive artifacts (shot noise) similar to normal micro-textural structures and disease manifestations, which significantly degrades the discriminative attributes in radiography images [3]. The signal-dependent characteristic of quantum noise has different influences at varying radiation levels (photon concentration). It is negatively correlated with the square root of the radiation exposure of the receptor plate and thus, quantum noise can be reduced by increasing the X-ray exposure time and radiation dose [4]. However, the prolonged exposure to high-dose X-ray radiation is hazardous to patients that may cause serious health issues, especially in the case of pediatric radiology [5]. Low dose X-ray imaging is a recent advancement in radiological imaging technique to prevent the hazards from exposure to high radiation level [5], but this technique suffers from amplified quantum noise and poor image quality due to the low photon flux density. Besides quantum noise, the electrical components and the structure of the films, intensifiers, or digital receptor plates generate a small amount of Gaussian or structural noise, respectively. However, with the advancement in medical imaging modalities, the effect of these noise has reduced [6], [7].

Image de-noising techniques are often preferred over using high dose X-ray radiation to subside the effect of quantum noise and to reduce health issues [8]. Quantum noise reduction in low-dose X-ray imaging has been widely researched, and various image restoration approaches in spatial as well as in frequency domain have been proposed over the past few years [2], [3], [9]. However, there is a trade-off between noise reduction and information loss (due to smoothing), which necessitates the selection of an efficient filtering algorithm. This paper aims to investigate the various benchmark filters to select an appropriate filter that effectively reduces the quantum noise while preserving important texture information. The quantitative and statistical evaluation of benchmark filters and its de-noising performance are investigated and compared based on disease classification accuracy using de-noised chest X-ray (CXR) images. This study ranks the filtering algorithm based on classification accuracy. Furthermore, the ability of the filters to preserve image quality without sacrificing the noise suppression is analyzed by two expert radiologists and based on traditional image quality metrics described in Table 4. The effectiveness of de-noising filters in terms of classification accuracy is validated with the ground truth data.

The next section presents an overview of techniques and the inclusion of signal-dependent quantum noise during the image acquisition process.

Quantum noise is signal-dependent and is predominant (since inherited from capturing device) in low contrast X-ray imaging modality, which gets added during image acquisition [10]. X-ray images are generated by passing the collimated X-ray beams through the patient body area to be imaged, and the attenuated beams are captured by a receptor plate on the other side, creating bone and soft-tissue shadow image as shown in Fig. 1(a). The random impinge (striking) of attenuated X-ray photons to the receptor plate for a given time interval causing uneven distribution of photons over the receptor’s surface (Fig. 1(b)) leads to irregular scene artifacts resulting in quantum noise.

The amount of quantum noise is inversely proportional to radiation exposure (photon concentration) and can be approximately expressed as exposurelevel the receptor. The noise decreases as the number of photons increases and if the number of photons used is quadrupled, the noise in the resultant image gets halved [11]. The random temporal scattering of photons over the receptor plate can be modeled by the standard Poisson distribution process [4], [12]. The probability distribution of photons captured on the receptor plate over a given time interval is described in Eq. (1).Px=e-λtλtkx!where, Px is the probability distribution of photons, λ is the expected number of photons, x is the measured number of photons and t is the given time interval.

The quantum noise follows the Poisson process where the expected photon count Ex is equal to the variance of photon count varx over given timeinterval as described in Eq. (2) which represents the signal dependent nature of noise. Moreover, the noise grows proportionately with the square root of the number of photons captured by the receptor plate, as described in Eq. (3).Ex=varx=λtstdx=λt

Fig. 2(a) represents a normal CXR image without noise and Fig. 2(b) represents a CXR image after the inclusion of synthetically generated Poisson noise using library function from MATLAB 2016a1. Since the quantum noise is not an additive or multiplicative type of noise, therefore the amount of noise included is entirely dependent on the input image pixel intensity. From the analysis of zoomed patch 1 and 2, as shown in Fig. 2(c) and 2(d), respectively, it is observed that the Poisson noise (or shot noise) creates a coarse appearance overlapped with the normal anatomical structure, which is challenging to analyze the micro-textural features on the CXR image accurately.

The significant contribution of this study are summarized as follows:

  • -

    Analyzed the performance of various benchmark filters for reducing quantum noise in CXR images.

  • -

    Implemented segmented lung ROI confined feature extraction technique on CXR images. Further classification accuracy is used as a measure to rank the performance of the de-noising filters.

  • -

    Quantitative, qualitative, subjective, and statistical analysis of different de-noising filters.

The rest of the paper is organized as follows. Section 2 presents a review of existing approaches for the elimination of quantum noise. In Section 3, the materials and methods used in this study are elaborated. Section 4 presents the quantitative, qualitative, and statistical analysis and their comparisons with benchmarked filters. The paper is concluded with some future directions in Section 5.

Section snippets

Review of quantum noise filtering approaches

This section elaborates on the quantum noise removal techniques used in this study. Most of the medical imaging modality suffers from noise [11], whose sources and characteristics are different. Filtering radiographic images reduces noise, restore corrupted texture information, and help to improve the expert’s diagnostic accuracy and robustness of automated CAD systems.

A filter is a squared kernel or mask used for smoothing noise by convolving the kernel window w(x,y) centering to each pixel

Materials and methods

In this section, we present the methodological design for experimentation and comparison of benchmark techniques. Fig. 3 depicts the overall roadmap of the methodology used in this study. All the experiments were implemented in MATLAB 2016a software. This study uses three evaluation methods (quantitative, qualitative, and statistical) to come up with an efficient quantum noise-reducing filter, which suppresses the noise while preserving the micro-textural pattern. The qualitative assessment was

Experimental results analysis and discussion

This section presents the comparative analysis of the various set of experimentation and results obtained from edge-aware quantum noise filtering techniques. The detailed discussion of the observed results is presented in the following subsections.

Conclusion

This study presented an extensive study of different benchmark filters, which are pervasively used for de-noising medical images. The objective was to introduce an efficient de-noising filter, which can preserve the diagnostic texture patterns of disease responses while effectively reducing the quantum noise in chest X-ray images. The tradeoff between texture preserving and de-noising performance of the filters were evaluated using quantitative, qualitative, subjective, and statistical

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgement

The authors would like to thank Dr. Satyabhuwan Singh Netam, Head, Department of Radiodiagnosis and Dr. Deepak Jain, Radiologist, Department of Radiodiagnosis, Pt. JawaharLal Nehru Memorial Medical College, Raipur (C.G.), India for their valuable guidance and suggestions.

References (98)

  • Q. Chen et al.

    Image denoising by bounded block matching and 3D filtering

    Signal Processing.

    (2010)
  • G. Easley et al.

    Sparse directional image representations using the discrete shearlet transform

    Appl. Comput. Harmon. Anal.

    (2008)
  • F. Shariaty et al.

    Application of CAD systems for the automatic detection of lung nodules

    Informatics Med. Unlocked.

    (2019)
  • M.F. Møller

    A scaled conjugate gradient algorithm for fast supervised learning

    Neural Networks.

    (1993)
  • D. Agnihotri et al.

    Variable Global Feature Selection Scheme for automatic classification of text documents

    Expert Syst. Appl.

    (2017)
  • F. Schopper et al.

    High resolution X-ray imaging with pnCCDs

    Nucl. Instruments Methods Phys. Res. Sect. A Accel. Spectrometers, Detect. Assoc. Equip.

    (2017)
  • A.J. Wright et al.

    Experimental animal endocarditis

    Mayo Clin. Proc.

    (1982)
  • S.W. Hasinoff

    Photon, Poisson Noise

    Comput. Vis. A Ref. Guid.

    (2014)
  • S. Lee et al.

    Poisson-gaussian noise analysis and estimation for low-dose X-ray images in the NSCT domain

    Sensors (Switzerland).

    (2018)
  • B. Goyal et al.

    Noise issues prevailing in various types of medical images

    Biomed. Pharmacol. J.

    (2018)
  • P. Gravel et al.

    A method for modeling noise in medical images

    IEEE Trans. Med. Imaging.

    (2004)
  • L. Wang et al.

    Noise removal for medical X-ray images in wavelet domain

    Electr. Eng. Japan.

    (2008)
  • M.T. Naseem et al.

    MRT letter: Quantum noise removal and classification of breast mammogram images

    Microsc. Res. Tech.

    (2012)
  • P. Sprawls, The Physical Principles of Medical Imaging, 2nd Ed., (2018). http://www.sprawls.org/ppmi2/NOISE/ (accessed...
  • B.K. Singh et al.

    Objective and optical evaluation of despeckle filters in breast ultrasound images

    IETE Tech. Rev.

    (2015)
  • H. Ibrahim et al.

    Simple adaptive median filter for the removal of impulse noise from highly corrupted images

    IEEE Trans. Consum. Electron.

    (2008)
  • U. Erkan et al.

    Different applied median filter in salt and pepper noise

    Comput. Electr. Eng.

    (2018)
  • V.S. Frost et al.

    A model for radar images and its application to adaptive digital filtering of multiplicative noise

    IEEE Trans. Pattern Anal. Mach. Intell.

    (1982)
  • M. Talha et al.

    Preprocessing digital breast mammograms using adaptive weighted frost filter

    Biomed. Res.

    (2016)
  • J. Jaybhay et al.

    A study of speckle noise reduction filters, signal image process

    An Int. J.

    (2015)
  • P. Perona et al.

    Scale-space and edge detection using anisotropic diffusion

    IEEE Trans. Pattern Anal. Mach. Intell.

    (1990)
  • M. Kumar et al.

    Jaya based functional link multilayer perceptron adaptive filter for Poisson noise suppression from X-ray images

    Multimed. Tools Appl.

    (2018)
  • E. Yilmaz, T. Kayikcioglu, S. Kayipmaz, Noise removal of CBCT images using an adaptive anisotropic diffusion filter,...
  • D. Bhonsle et al.

    Medical image denoising using bilateral filter

    Int. J. Image, Graph. Signal Process.

    (2012)
  • Zhiwu Liao et al.

    Medical image blind denoising using context bilateral filter

  • M.D.G. Raul

    Guided filter technique: various aspects in image processing

    Int. J. Recent Innov. Trends Comput. Commun.

    (2016)
  • K.N. Chaudhury, K. Rithwik, Image denoising using optimally weighted bilateral filters: A sure and fast approach, in:...
  • Buyue Zhang et al.

    Adaptive bilateral filter for sharpness enhancement and noise removal

    IEEE Trans. Image Process.

    (2008)
  • K. He et al.

    Guided image filtering

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2012)
  • A. Hosni et al.

    Fast cost-volume filtering for visual correspondence and beyond

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2013)
  • H.K. Westin et al.

    Adaptive image filtering

    (2000)
  • M.A. Abd El-Fattah et al.

    Speech enhancement using an adaptive wiener filtering approach

    Prog. Electromagn. Res.

    (2008)
  • K. Dabov et al.

    Image denoising by sparse 3-D transform-domain collaborative filtering

    IEEE Trans. Image Process.

    (2007)
  • P.R. Vn

    Denoising of magnetic resonance and X-Ray images using variance stabilization and patch based algorithms

    Int. J. Multimed. Its Appl.

    (2012)
  • T. Kirti et al.

    Poisson noise reduction from X-ray images by region classification and response median filtering

    Sadhana.

    (2017)
  • A. Zhang et al.

    A Shearlet-based algorithm for quantum noise removal in low-dose CT images

    Med. Imaging 2016 Image Process.

    (2016)
  • D.D. Muresan, T.W. Parks, Adaptive principal components and image denoising., in: ICIP, 2003: pp....
  • J. Salmon et al.

    Poisson noise reduction with non-local PCA

    J. Math. Imaging Vis.

    (2014)
  • G. Treece

    The bitonic filter: linear filtering in an edge-preserving morphological framework

    IEEE Trans. Image Process.

    (2016)
  • Cited by (33)

    • An optimized and efficient android malware detection framework for future sustainable computing

      2022, Sustainable Energy Technologies and Assessments
      Citation Excerpt :

      However, the supervised algorithms that train single learners usually suffer from statistical, computational, and representation problems [43]. Further, we also evaluated and compared the performance of other benchmark classification methods, namely—Fine Tree (FT), Kernel Naïve Bayes (kNB), Support Vector Machine with the linear kernel (SVM-L), quadratic kernel (SVM-Q), and cubic kernel (SVM-C) [44] and weighted K-Nearest Neighbour (wKNN) has been considered. This paper used a publicly available android malware dataset (CICInvesAndMal2019) [23], which is an extended version of CICAndMal2017 [24].

    • Disease Localization and Severity Assessment in Chest X-Ray Images using Multi-Stage Superpixels Classification

      2022, Computer Methods and Programs in Biomedicine
      Citation Excerpt :

      Further, the choice of the classification algorithms is inspired by the fact that these algorithms can be expeditiously trained the models using limited training data without much impairing the performance. Moreover, they are pervasively used in literature to classify the medical images, which motivated us to use them in our study [5,62,78]. Subsequently, an infection map is generated by overlaying the heatmap images for infected superpixels and soft-tissue regions on the input CXR image, as shown in Fig. 9.

    • A novel method based on Wiener filter for denoising Poisson noise from medical X-Ray images

      2022, Biomedical Signal Processing and Control
      Citation Excerpt :

      Optimization algorithms, an artificial intelligence application, were used for 2D FIR filter design. In the current study of Chandra et al., filtered images were classified using different filter structures by the ANN classifier [11]. Taking this study as a reference, the Wiener filter, which is the most successful on average, was used in this study.

    View all citing articles on Scopus
    View full text