An efficient and robust approach for biomedical image retrieval using Zernike moments
Introduction
In this era with ever increasing number of hospitals and medical institutions, the number of biomedical images related to patient diagnosis is growing drastically. Efficient utilization of this database for search and retrieval of medical images is very important for accurate patient diagnosis and for any further R&D (research and development) in medical sciences. To meet this requirement, content-based image retrieval (CBIR) systems were introduced.
CBIR systems search and retrieve images based on the visual contents of an image such as color, texture, shape, and spatial layout, etc. Such sort of information about an image is stored in the form of image features (local or global). Global features represent the visual features of an entire image as whole, whereas the local features represent the visual features of regions or objects to describe an image. Accuracy of any retrieval system highly depends on the type of features used and their capability to distinguish among different image classes/groups. The detailed survey about various feature extraction techniques employed in CBIR systems can be found in [1], [2], [3], [4], [5].
Since, medical images exists in various formats, in this paper, the retrieval of only computed tomography (CT) and magnetic resonance imaging (MRI) images is performed. Initial works for retrieval of CT and MRI images can be seen in [6]. They have used bit plane histogram and hierarchical bit plane histogram along with the cumulative distribution functions for this purpose. In [7], a system for classification of benign and malignant breast masses based on texture and shape features in sonography images is proposed. Quellec et al. [8] proposed a medical retrieval system for diabetic retinopathy and mammographic databases. They used an optimized wavelet transform by adapting the wavelet basis within the lifting scheme framework and assign weights for each wavelet sub-band. Another work based on wavelet transform for brain image retrieval can be seen in [9]. The co-occurrence matrix based retrieval of CT and MRI images is proposed in [10]. A fast and robust local-structure-based region-of-interest (ROI) retrieval system for brain MRI images is given in [11]. Soreson et al. [12] provided a quantitative analysis of pulmonary emphysema in CT images based on LBP (local binary pattern). They have collaboratively used the LBP, joint LBP and intensity histograms for characterizing ROIs. Later, Peng et al. [13] developed a uniformity estimation method (UEM) for local brightness and structures to discover the pathological changes in the chest CT images. They extract features by proposing an extension of rotation invariant LBP and orientation difference of gradient vector to represent the brightness and structure in the image. Quddus et al. [14] proposed a method for retrieval of 2D MRI images in multimodal and noisy observations. Results for retrieval of 2D MRI brain images are better in terms of accuracy, speed, robustness and multimodality. Among the recent works for the retrieval of CT and MRI images includes the work of Murala and Wu [15], [16], [17], [18]. Murala and Wu in [15] proposed a new feature extraction technique called local co-occurrence ternary patterns (LTCoP) which encodes the co-occurrence of similar ternary edges in an image. Given a central pixel, the value of LTCoP is calculated based on the first order derivatives in eight directions. The reported results show the superiority of their proposed approach in comparison to techniques like LBP, LDP (local directional pattern), LTP (local ternary pattern) etc. Later on, authors proposed another novel feature descriptor for biomedical image retrieval called local mesh patterns (LMeP) [16]. LMeP encodes the relationship among the surrounding neighbors for a given reference pixel in an image, contrary to LBP that encodes the relationship between the given reference pixel and its surrounding neighbors. This approach was further extended by authors in [17] to derive LMePVEP i.e. local mesh peak valley edge patterns. It was shown that LMePVEP demonstrated a significant improvement in terms of ARR and ARP as compared to LBP and LBP variants. Following the same trend, Dubey et al. in [19] proposed a technique called local diagonal extrema pattern (LDEP) for CT image retrieval. LDEP uses first order local diagonal derivatives to determine the values and indexes of the local diagonal extremas. These indexes are then used to form a feature descriptor for an entire image. To demonstrate the effectiveness and superiority of their approach, authors compared their results with those obtained by other state-of-the-art methods such as LMeP, LTCoP, LTP, CSLBP, and LBP etc. and shown improved retrieval accuracy at reduced computational time. The biggest advantage offered by LDEP over other existing methods is its reduced dimensionality of feature vector, thus, offering higher performance using very less number of features.
However, like LBP and its multiple variants, LDEP is also a local feature extractor. While working on medical images of different body parts, the gross view of an image is very crucial for accurate identification, which the local features fail to capture. Therefore, due to the high dimensionality of feature vectors and lack of representation capability of local features extracted by LBP [20] and its different variants, we shifted our focus towards finding a global feature descriptor that can efficiently describe and represent image features (both shape and texture) for biomedical image retrieval purpose.
Among the numerous global feature descriptors existing in the literature, Zernike moments (ZMs) [21] are found to be very suitable for biomedical image retrieval application because of its useful characteristics such as excellent image representation capability, rotation (and scale) invariance, noise robustness, small feature size etc. All these properties of ZMs made it highly successful in various image processing and pattern recognition applications [22], [23], [24], [25], [26], [27], [28], [29], [30], [31], [32], [33], [34], [35], [36], [37], [38], [39], [40], [41], [42], [43] including biomedical image enhancement/segmentation [43].
In this paper, we propose the use of ZMs for the retrieval of CT and MRI images and compare its performance with standard benchmark approaches like LBP, ULBP and the recently published state-of-the-art LDEP method on two standard test databases namely Emphysema-CT and OASIS-MRI database. Due to the superior performance of LDEP approach over the other existing methods like LTP, LTCoP, LMeP etc. as shown in [19], we therefore, compared our proposed ZMs-based approach with LDEP only. Additional experiments have been performed to test the effectiveness of the proposed ZMs-based approach under noisy conditions and to demonstrate its superiority (in terms of retrieval performance) over the existing and recently published approaches of biomedical image retrieval.
The rest of the paper is organized as follows. Section 2 describes various existing feature extraction methods for CT and MRI image retrieval. Section 3 describes the proposed ZMs-based system for retrieval of CT and MRI images. Section 4, presents the experimental set up and results obtained after using the proposed and the existing feature descriptors on two standard medical databases. Also, a comparative performance analysis of the proposed approach with recently published state-of-the-art approaches for biomedical image retrieval is presented. Section 5, presents the computational complexity of the proposed approach and all other compared feature descriptors for biomedical image retrieval application. Finally, conclusion is presented in Section 6.
Section snippets
Feature extraction methods
Feature extraction plays a very crucial role in the efficient working of any retrieval system both in terms of the retrieval accuracy as well as the computation time taken to achieve that accuracy. The goal of feature extraction part is to extract such features that can strongly represent an image using only little number of them. The local descriptors such as LBP, ULBP and LDEP are some of the well-known feature extraction techniques that have been used for retrieval of CT and MRI images.
Proposed Zernike moments (ZMs)-based image retrieval system
All the feature descriptors described in earlier section are local in nature. Biomedical retrieval system involves medical images of various body parts; therefore, a global aspect of these images is very crucial for a clear distinction among them. Since, the local features are unable to capture the gross aspect of an image; we propose the use of Zernike moments (ZMs) as a global descriptor for the retrieval of medical images.
Experiments and results
In this section, we investigate the performance of the proposed ZMs-based approach for CT and MRI image retrieval and compare its performance with existing approaches such as LDEP, ULBP and LBP. All algorithms are implemented in MATLAB R2010b under Microsoft Windows 7 environment on a PC with 2.13 GHz CPU and 3 GB RAM. All experiments have been conducted on two standard medical image databases namely Emphysema-CT database [55] for CT image retrieval and OASIS-MRI database [56] for MRI image
Discussion
The results presented in Section 4.1 to 4.5 demonstrate the superiority of the proposed ZMs-based approach over the state-of-the-art retrieval approaches in its ability to correctly retrieve CT and MRI images. However, there are few scenarios where the proposed ZMs-based approach sometimes fails to retrieve correct images. The following depicts possible reasons for such behavior of the proposed approach.
- a)
As mentioned in Section 3, we have used outer circle mapping approach for the computation of
Computational complexity
The computational time associated with a feature descriptor is as equally important as its retrieval performance. This computational time can be divided into phases: 1) Feature Extraction Time: Time spent in extracting features from the query and database images, and 2) Retrieval (Classification) Time: Time spent in comparing the features of the query image with the database images. Speed of extracting similar images to the query image depends on the feature vector length of the feature
Conclusion
In this paper, an effort has been made to explore the performance of Zernike moments (ZMs) for biomedical image retrieval application. ZMs are orthogonal moments and belong to a category of global feature descriptors. ZMs extract gross information from an image which the local feature extractors like LBP, LDEP etc. fail to capture. This gross view is very crucial for accurate distinction between medical images of different body parts. Also, their property of noise insensitivity makes them
References (59)
- et al.
Image retrieval: current techniques, promising directions and open issues
J. Vis. Commun. Image Representation
(1999) - et al.
A survey of content-based image retrieval with high-level semantics
Pattern Recogn.
(2007) - et al.
A review of content-based image retrieval systems in medical applications—Clinical benefits and future directions
Int. J. Med. Inf.
(2004) - et al.
Wavelet optimization for content-based image retrieval in medical databases
J. Med. Image Anal.
(2010) - et al.
Texture feature extraction based on uniformity estimation method for local brightness and structure in chest CT images
Comput. Biol. Med.
(2010) - et al.
Local ternary co-occurrence patterns: a new feature descriptor for MRI and CT image retrieval
Neurocomputing
(2013) - et al.
MRI and CT image indexing and retrieval using local mesh peak valley edge patterns
Signal Process. Image Commun.
(2014) - et al.
Spherical symmetric 3D local ternary patterns for natural, texture and biomedical image indexing and retrieval
Neurocomputing
(2015) - et al.
A comparative study of texture measures with classification based on feature distributions
Pattern Recogn.
(1996) - et al.
Invariant character recognition with Zernike and orthogonal Fourier–Mellin moments
Pattern Recogn.
(2002)