An integrated scattering feature with application to medical image retrieval

https://doi.org/10.1016/j.compeleceng.2018.01.027Get rights and content

Abstract

Scattering transform has been successfully applied to medical image retrieval because it provides extensive semantic representations of an image. To efficiently handle the scattering transform results, the coefficient matrices are usually compressed to vectors. However, the existing features derived from the compressed vectors only consider one distribution information of the original image. To address this problem, this paper proposes an integrated scattering feature for medical image retrieval. The proposed method integrates two types of compressed scattering data from different perspectives, namely data concentration and canonical correlation analysis (CCA). For each integration model, we also give a corresponding feature representation strategy that takes account of more comprehensive characteristics of original medical image. Experiments on two benchmark medical computed tomography (CT) image databases demonstrate the superiorities of the proposed features over several state-of-the-art methods.

Introduction

Medical image retrieval (MIR) is an important topic in the fields of medical image analysis and pattern recognition owning to its extensive applications. The performance of a MIR system heavily depends on the used features. Drawing on the experience of general multimedia retrieval techniques [1], [2], [3], [4], [5], a great deal of algorithms have been developed to extract features of medical images. Because there are many texture-like regions on medical images, the existing features for traditional texture image representation are popularly used in several MIR systems. The well-known local binary pattern (LBP) [6], [7] and its improvements [8], [9], [10], [11], [12], [13] have achieved promising performance. The key idea of these methods is to develop different local encoding strategies to describe the local image contents from various perspectives.

Inspired by the success of deep learning techniques [14], the scattering transform [15], a variation of deep convolutional networks, has been used to derive features for MIR. In [16], the authors proposed a histogram of compressed scattering coefficients (HCSC) feature from medical images. HCSC first employs a projection operation to compress the scattering coefficient matrix to a vector by a given direction, and then describes the compressed vectors via the bag-of-words (BoW) model. Though HCSC has achieved satisfactory performance, it has following limitations. First, HCSC only considers one projection direction for compression such that just partial information of the image is used. Fig. 1 illustrates the vertical and horizontal projections of a scattering coefficient matrix. It can be seen that they contain different intensity distribution characteristics of the matrix, but only one of them is applied in HCSC. It is necessary to take account of all these information to derive features. Second, the derivation of codebook used in BoW model is time consuming because of the clustering operation, which adversely affects the efficiency of HCSC.

To address the above issues, we develop an integrated scattering feature for medical image retrieval in this paper. Rather than only considering one intensity distribution information of a scattering coefficient matrix, we first integrate two different intensity distribution information from following perspectives, namely data concentration and canonical correlation analysis (CCA) [17]. Two corresponding feature representations are also given, which include more comprehensive distribution information of the original image in contrast with some related features. The developed scheme is evaluated on two benchmark medical computed tomography (CT) image databases, and the comparison results demonstrate that it outperforms several state-of-the-art methods.

The remainder of this paper is organized as follows. Section 2 describes the proposed algorithm in detail, and Section 3 illustrates several comparison results obtained by the proposed algorithm. Section 4 finally presents some conclusions and future work.

Section snippets

Proposed approach

In this section, we will detail the proposed feature extraction approach for medical images. Considering a MIR system, let A denotes a collection of all training images. The proposed method consists of following modules, i.e., scattering coefficient extraction, scattering coefficient integration, and scattering feature representation, respectively.

Experimental results

Several comparison results are reported in this section to evaluate the performance of the proposed algorithm. We first introduce the setting of all experiments, and then present the retrieval results of all competing algorithms. A comparison of time complexity is finally given.

Conclusion

In this paper, we have presented an integrated scattering feature for medical image retrieval. The proposed feature takes account of different scattering information that is able to provide high level representations of the original medical image contents. Compared with the existing HCSC feature, which is derived from one type of compressed scattering coefficients, the proposed one further makes use of data integration to consider two different compressed data. Hence, the proposed feature

Acknowledgments

This work was partially supported in part by the National Natural Science Foundation of China (Nos. 61702129, 61772149, and 61320106008), Guangxi Colleges and Universities Key Laboratory of Intelligent Processing of Computer Images and Graphics (No. GIIP201703), and Guangxi Key Research and Development Program (Nos. AB17195057 and AB17195025).

Rushi Lan is an assistant professor of the School of Computer Science and Information Security, Guilin University of Electronic Technology, China. His current research interests include medical image processing, pattern classification, and machine learning.

References (25)

  • L. Liu et al.

    Weighted joint sparse representation for removing mixed noise in image

    IEEE Trans Cybern

    (2017)
  • T. Ojala et al.

    Multiresolution gray-scale and rotation invariant texture classification with local binary patterns

    IEEE Trans Pattern Anal Mach Intell

    (2002)
  • Cited by (12)

    • Fine-grained correlation analysis for medical image retrieval

      2021, Computers and Electrical Engineering
      Citation Excerpt :

      The CCA-based feature fusion method has been widely applied in face recognition. Ref. [15] also proposed an integrated scattering feature, called scattering transform canonical correlation analysis (ST-CCA), for medical image retrieval. It integrated two types of compressed scattering data via canonical correlation analysis and took account of more comprehensive characteristics of medical images.

    • Complex background classification network: A deep learning method for urban images classification

      2020, Computers and Electrical Engineering
      Citation Excerpt :

      Experiments are conducted to demonstrate the performance of proposed method. Our network draws on the idea that image classification to predict all information may depend on a key object [14]. The main contribution of our work is twofold: (i) We propose a novel Complex Background Classification Network (CBCNet) for urban image classification, in which a detection sub-network is developed to recognize key information in urban images and an evaluation sub-network is developed to optimize object parameters and improve the robustness of the whole network. (

    View all citing articles on Scopus

    Rushi Lan is an assistant professor of the School of Computer Science and Information Security, Guilin University of Electronic Technology, China. His current research interests include medical image processing, pattern classification, and machine learning.

    Huadeng Wang is an associate professor of the School of Computer Science and Information Security, Guilin University of Electronic Technology, China. His current research interests include medical data mining, machine learning, and computer graphics.

    Si Zhong is a lecture of the School of Computer Science and Information Security, Guilin University of Electronic Technology, China. Her current research interests include image processing and information retrieval.

    Zhenbing Liu is a professor of the School of Computer Science and Information Security, Guilin University of Electronic Technology, China. His current research interests include pattern recognition, image processing, and machine learning.

    Xiaonan Luo is a professor of the School of Computer Science and Information Security, Guilin University of Electronic Technology, China. He is an IEEE senior member. His current research interests include computer graphics, medical image processing, and location based services.

    Reviews processed and recommended for publication to the Editor-in-Chief by Associate Editor Dr. Huimin Lu.

    View full text