Medical image classification using spatial adjacent histogram based on adaptive local binary patterns
Introduction
Medical images have played an important role in the diagnostic workup of patients. Automated classification of medical images is a desirable tool to assign the interpretation of images, and then would help the expert in diagnosis of diseases [1], [2], [3]. Compared with general image recognition, medical image recognition is more challenging because of the higher ambiguity and complexity; most of the medical image contents are quite similar, but also different in their emphasis.
In terms of the features used for medical image recognition, it can be mainly classified into three groups: shape, color, and texture features. For example, in [4], shape features such as moment invariants and Fourier descriptor are employed to classify medical X-ray images. A color vector field is considered in [5] for improving the performance of endoscopic image classification.
The local binary patterns, first proposed by [6], are widely considered as a state-of-the-art image feature descriptor among texture descriptors, since it can more effectively describe texture information. It has been successfully applied to many applications, such as face recognition, texture classification, scene recognition, human detection and others. LBP has several attractive advantages: it has proven to be a powerful discriminator with low computational cost, it is robust against changes in image intensity, and it can be easily implemented. Due to these merits, it makes a good choice for extracting fine features for medical images.
However, the standard LBP still suffers from several drawbacks, including limited semantic description of local patterns, sensitive to non-uniform patterns and affine transformation, and missing of efficient spatial encoding among patterns. To overcome these shortcomings, numerous works [9], [10], [11] focused on improving LBP in recent years, in terms of rotation-invariant, multi-scale, the utilization of non-uniform patterns, and so on. There are two types of LBP patterns: uniform and non-uniform patterns. Some works, such as [7], only considered uniform patterns for extracting LBP features since non-uniform patterns involve noise and high dimensionality. And the work [8] proposed a hierarchical multiscale LBP to further utilize the information of non-uniform patterns. They also certify that, the percentage of non-uniform patterns increases as the neighborhood radius increases. To reduce the LBP dimensionality, center-symmetric local binary patterns (CSLBP) [15] is studied and applied to image recognition. Since LBP is sensitive to noise in near uniform regions, Local Ternary Pattern [14] with three value coding scheme was proposed to address this problem. The rotation invariant [6] descriptor can be obtained through the circular neighborhood definition, but in some cases the anisotropic structural information is lost. To utilize these anisotropic structural information, a novel elliptical binary pattern (EBP) [16] has been proposed for face recognition, in which elliptical neighborhood definitions are studied. Completed LBP [13] utilizes both the sign and magnitude information in the difference between the central pixel and the neighborhood pixels. In the work [18], LBP is combined with Gabor filters to achieve a better classification performance. The study [19] extracted the most frequent patterns in LBP histogram and formed a novel descriptor, which achieved better performance with this technique. Mesh-LBP [21] is novel method which computed the mesh-local binary pattern on a triangular-mesh manifold. In [22], a scale- and rotation-invariant LBP is proposed, in which the rotation-invariant is combined with a scale-adaptive texton for texture classification. SOALBP [23] constructed a novel scale- and orientation invariant LBP feature combined in a multi-resolution representation, which has been proven superior in texture classification.
The basic idea behind LBP is that it describes an image by local patterns. The existing methods have been proven to improve the LBP to some extent by reconfiguring or utilizing the patterns. However, most of existing works encode the binary patterns in a fixed neighborhood radius. This fixed neighborhood radius strategy is irrelevant to local image content and disregards micro-structure information of the multi-scale patterns. Intuitively, the micro-structures, i.e. the spatial relationships among local patterns generated by adaptive radius, provide crucial feedback in disambiguating texture information especially for complex medical images, i.e. microscope images that involve with pathological changes. This subsequently leads to improved recognition performance. To this end, our target is to design a novel LBP histogram representation for medical images to (1) compute the local binary patterns in an adaptive neighborhood radius, and (2) encode micro-structures among the multi-scale patterns. In the first stage, with the help of gradient operators, we obtain a gradient map from each original image, and the adaptive LBP neighborhood radius could be then determined for each pixel by utilizing the gradient information. As a result, our adaptive strategy will assign a relatively small radius to pixels that are located in local regions with dramatic gray variation, while assigning a relatively large radius to pixels that are located in local regions with slight gray variation. This adaptive technique will provide the image with rich micro-structure textures, which is discriminative in image representation. Then in the next stage, we propose a spatial adjacent histogram based on adaptive LBP radius to describe these discriminative micro-structure features. Finally, the adaptive LBP radius and spatial adjacent histogram strategies produce a much more powerful LBP variant, which performs well in four benchmark medical datasets and compares favorably to other methods.
In this context, our contribution is threefold.
- 1)
Using the adaptive strategy we proposed, the neighborhood radius of LBP is determined based on local image content, therefore more adaptive and useful features can be obtained.
- 2)
We propose to use spatial adjacent histogram to encode the micro-structures produced by adaptive strategy, which results in convincing improvement on the standard LBP histogram.
- 3)
Our approach also considers three LBP coding schemes, i.e. set the threshold T in three different ways when computing the LBP value. And we further evaluate the three LBP coding schemes in order to find which one performs more competitive in medical image classification.
The remainder of this paper is organized as follows. In Section 2, we introduce the proposed algorithm, i.e. spatial adjacent histogram based on adaptive local binary patterns for image classification. Section 3 presents an extensive set of experimental evaluations on four medical image datasets, and finally, in Section 4, we draw the conclusions.
Section snippets
Spatial adjacent histogram based on adaptive local binary patterns for image classification
In this section, we propose a novel idea using spatial adjacent histogram based on adaptive local binary patterns for medical image classification. We first present a concise review of standard local binary patterns (LBP) in subsection 2.1. Next, we explain how to determine the adaptive radius for each pixel in subsection 2.2, then in subsection 2.3, three coding schemes are introduced to compute the LBP value for each pixel. In subsection 2.4, we propose a spatial adjacent histogram technique
Experimental datasets
The medical datasets for the evaluation of the proposed method are selected based on their variety and broadness of their use. Experiments are performed on the following four reference datasets.
Conclusion and future work
In this paper, we proposed a novel framework based on our adaptive local binary patterns and spatial adjacent histogram for medical image classification. We first employed gradient operator to determine the adaptive neighborhood radius of each pixel. Then three coding schemes based on adaptive radius were introduced for processing the LBP histograms. To capture the discriminative micro-structures features produced by the adaptive radius, we proposed to using spatial adjacent histogram strategy
Conflict of interest statement
None declared.
Acknowledgments
This work was supported by the National Natural Science Foundation of China (61472161, 61133011, 61402195, 61502198, 61303132, 61202308), Science and Technology Development Project of Jilin Province (20140101201JC), the Science and Technology Plan Project of Wenzhou of China (G20140048) and the Program of China Scholarships Council (No. 201406170116).
References (35)
Computer-aided diagnosis in medical imaging: historical review, current status and future potential
Comput. Med. Imaging Graph.
(2007)- et al.
Content-based medical image classification using a new hierarchical merging scheme
Comput. Med. Imaging Graph.
(2008) - et al.
A novel extended local binary pattern operator for texture analysis
Inf. Sci.
(2008) - et al.
Survey on LBP based texture descriptors for image classification
Expert Syst. Appl.
(2012) - et al.
Building global image features for scene regognition
Pattern Recognit.
(2012) - et al.
A scale-and orientation-adaptive extension of Local Binary Patterns for texture classification
Pattern Recognit.
(2015) - et al.
Automatic Phases Recognition in Pituitary Surgeries by Microscope Images Classification. In Information Processing in Computer-assisted Interventions
(2010) - et al.
Statistical models of appearance for medical image analysis and computer vision. Medical imaging 2001
Int. Soc. Opt. Photon.
(2001) - M. Häfnera, M. Liedlgruber, A. Uhl, A. Vécsei, F. Wrba, Color treatment in endoscopic image classification using...
- et al.
Multiresolution gray-scale and rotation invariant texture classification with local binary pattern
IEEE Trans. Pattern Anal. Mach. Intell.
(2002)
Noise-resistant local binary pattern with an embedded error-correction mechanism
IEEE Trans. Image Process.
Robust image region descriptor using local derivative ordinal binary pattern
J. Electron. Imaging
A completed modeling of local binary pattern operator for texture classification
IEEE Trans. Image Process.
Enhanced local texture feature sets for face recognition under difficult lighting conditions
IEEE Trans. Image Process.
Cited by (85)
Efficient unsupervised learning of biological images with compressed deep features
2023, Image and Vision ComputingSine–Cosine-Barnacles Algorithm Optimizer with disruption operator for global optimization and automatic data clustering
2022, Expert Systems with ApplicationsMagnetoelectric Actuating Device for Providing Delicate Capture of the Object of Manipulation in Intelligent Robots
2024, Lecture Notes in Networks and SystemsLMFD: lightweight multi-feature descriptors for image stitching
2023, Scientific ReportsBiomedical image classification based on a feature concatenation and ensemble of deep CNNs
2023, Journal of Ambient Intelligence and Humanized Computing