An SVM-based distal lung image classification using texture descriptors

https://doi.org/10.1016/j.compmedimag.2011.11.001Get rights and content

Abstract

A novel imaging technique can now provide microscopic images of the distal lung in vivo, for which quantitative analysis tools need to be developed. In this paper, we present an image classification system that is able to discriminate between normal and pathological images. Different feature spaces for discrimination are investigated and evaluated using a support vector machine. Best classification rates reach up to 90% and 95% on non-smoker and smoker groups, respectively. A feature selection process is also implemented, that allows us to gain some insight about these images. Whereas further tests on extended databases are needed, these first results indicate that efficient computer based automated classification of normal vs. pathological images of the distal lung is feasible.

Introduction

The lungs are the essential respiratory organs. They are divided into two anatomic and functional regions: the air conduction system (that includes the trachea, bronchi, and bronchioles) and the gas-exchange region made of alveolar sacs. Whereas the conduction airways can be explored during bronchoscopy, the alveolar region is currently investigated only in vitro, using invasive techniques such as open lung biopsies. Recently, a new endoscopic technique, called fibered confocal fluorescence microscopy (FCFM), has been developed that enables the visualization of the more distal regions of the lungs in vivo [1]. The technique is based on the principle of fluorescence confocal microscopy, where the microscope objective is replaced by a fiberoptic miniprobe, made of thousands of fiber cores. The miniprobe can be introduced into the 2 mm working channel of a flexible bronchoscope to produce in vivo endomicroscopic imaging of the human respiratory tract in real time. This very promising technique could replace lung biopsy in the future and might prove to be helpful in a large variety of diseases, including interstitial lung diseases [2].

Images acquired with FCFM represent the alveolar structure (Fig. 1, Fig. 2), which can be altered by distal lung pathologies. Medical experts are still examining visually these images to search for typical properties that could be of interest for discriminating pathological subjects from healthy ones. Our aim is thus to provide the clinician with a computer aided-diagnosis tool, so as to help him to analyze these images, by automatically classifying FCFM images into healthy or pathological ones. The most crucial step when designing an automatic image classification system is the choice of relevant features to describe the image. In a preliminary work [7], we have designed a first classification system, so as to discriminate healthy cases from pathological cases, in which images are described by an ad hoc feature vector based on a visual analysis of the images. But several generic, low-level texture descriptions have been proved very efficient for various image description tasks [8]. We thus investigate in this paper those state-of-the-art feature descriptors to describe our FCFM images and compare their results to the ones obtained with the ad hoc feature vector so as to determine which descriptors are best adapted to the classification of FCFM images. Using a support vector machine (SVM) for classification, we show that one low-level feature vector is better adapted to the discrimination between pathological and healthy images. Furthermore, we point out that some insight about these images can be gained by studying, through a feature selection process, the most relevant features.

The remainder of this paper is thus organized as follows: the imaging technique is detailed in Section 2, our classification system is described in Section 3, and results and discussion are provided in Section 4. Conclusion is presented in Section 5 along with perspectives for this work.

Section snippets

FCFM images

FCFM images represent the alveolar structure, made of elastin fiber (Fig. 1, Fig. 2, Fig. 3, Fig. 4), with a lateral resolution of 3.5 μm. The FCFM imaging technique enables to observe the elastin as it is the main endogenous fluorophore at 488 nm in the respiratory tract from the bronchus to the alveoli. The elastin framework appears as a network of (almost) continuous lines in healthy subjects, as shown in Fig. 1. In [2], morphometric analysis of the alveolar structures was performed, based on

Feature extraction

In image classification system, feature extraction is still a critical step. Features are usually adapted to the image content. For instance, blood cell images can be described by frequency analysis, using Fourier–Mellin transform [3]; wooden image texture and color can be described by LBP (local binary patterns) descriptor [4]; protein sub-cellular images can be described by their texture, with cooccurrence matrix using Haralick statistical parameters [5]. The image description can also be

Experimental protocol

The SVM classifier and SVM-RFE based feature selection [16] are implemented using the SVM and kernel methods Matlab toolbox [17]. The number of images for the smoker and the non-smoker groups is provided in Table 3. Because of the relatively small number of non-smoker and smoker images, a 10 fold cross-validation process is used. The system performance is assessed with the mean recognition rates ± standard deviation on the 10 folds.

Classification results

From the recognition rates reported in Table 4, one can see

Conclusion

The present work deals with the classification of a new category of images of the distal lung. The images were acquired using fibered confocal fluorescence microscopy, a technique that enables the observation of in vivo alveolar structures. Such images are not well described so far and are difficult to discriminate by pathologists and respiratory physicians. In this paper, we have proposed several possible descriptions for these images, one based on their visual analysis, and the other ones

References (17)

  • ThibervilleL. et al.

    In vivo imaging of the bronchial wall microstructure using fibered confocal fluorescence microscopy

    Am J Respir Crit Care Med

    (2007)
  • ThibervilleL. et al.

    Human in vivo fluorescence microimaging of the alveolar ducts and sacs during bronchoscopy

    Eur Respir J

    (2009)
  • DahmenJ. et al.

    Automatic classification of red blood cells using Gaussian mixture densities

  • OjalaT. et al.

    Gray scale and rotation invariant texture classification with local binary patterns

    ECCV

    (2000)
  • HaralickR.M. et al.

    Textural features for image classification

    Syst Man Cybernet

    (1973)
  • MikolajczykK. et al.

    Scale and affine invariant interest point detectors

    Int J Comput Vision

    (2004)
  • PetitjeanC. et al.

    Classification of in-vivo endomicroscopic images of the alveolar respiratory system

  • MikolajczykK. et al.

    A performance evaluation of local descriptors

    IEEE Trans Pattern Anal Mach Intell

    (2005)
There are more references available in the full text version of this article.

Cited by (22)

  • Multiclass malware classification via first- and second-order texture statistics

    2020, Computers and Security
    Citation Excerpt :

    The LBP descriptor computes the distribution of binary patterns in the circular neighbourhood (characterized by radius R and number of neighbours P) of each pixel. The best-performing LBP feature vectors for a set of values (R = 1, P = 8), (R = 2, P = 16), and (R = 3, P = 24) followed by SVM classification (Désir et al., 2012) resulted in 94.14% and 94.03% accuracy for the Malimg dataset and Dataset 2, respectively. Chen et al. (2016) used the multi-scale completed LBP (MS-CLBP) descriptor, providing two implementations.

  • Automatic detection of body packing in abdominal X-ray images

    2020, Forensic Imaging
    Citation Excerpt :

    Therefore, a feature extraction method was proposed to reduce the influence of local noise and the size of the feature vectors. An SVM[17] classifier was implemented, which was based on statistical learning theory and had demonstrated optimal performance for small samples and nonlinear classifications [18–20]. To balance the classification quality and calculation rate, the radial basis kernel (RBF) and linear kernel functions [21] were selected.

  • Image computing for fibre-bundle endomicroscopy: A review

    2020, Medical Image Analysis
    Citation Excerpt :

    Classification of frames on pre-determined, clinically defined cohorts based on their content is currently the most investigated area of FBEµ image computing research. An abundance of studies have applied binary as well as multi-class classification on endomicroscopic images of a range of organ systems in an attempt to identify cancer in ovarian epithelium (Srivastava et al., 2005,2008), abnormalities in distal lung alveolar structures (Desir et al., 2012a; Désir et al., 2010,2012b; Hebert et al., 2012; Heutte et al., 2016; Koujan et al., 2018; Petitjean et al., 2009; Saint-Réquier et al., 2009), informative frames in brain (Izadyyazdanabadi et al., 2017; Izadyyazdanabadi et al., 2018) and pulmonary videos (Leonovych et al., 2018; Perperidis et al., 2016), cancerous nodules in the airways (Gil et al., 2017; He et al., 2012; Rakotomamonjy et al., 2014) and distal lung (Seth et al., 2016), pathological epithelium in the oropharyngeal cavity (Aubreville et al., 2017; Jaremenko et al., 2015; Vo et al., 2017), changes in oesophageal epithelium in cases of Barrett's oesophagus (Ghatwary et al., 2017; Hong et al., 2017; Veronese et al., 2013; Wu et al., 2017), adenocarcinoma (Ştefănescu et al., 2016), colorectal polyps (André et al., 2012b; Zubiolo et al., 2014) and celiac disease (Boschetto et al., 2016a) in intestinal epithelium, neoplastic tissue in breast nodules (Gu et al., 2017), as well as two types of common brain tumours, glioblastoma and meningioma (Kamen et al., 2016; Murthy et al., 2017; Wan et al., 2015). Methodologically, most of the aforementioned studies employed the same basic structure, defining a hand-crafted feature space descriptive of the underlying imaged structure, training a range of classifiers, to distinguish between pre-determined frame categories.

  • Scattering features for lung cancer detection in fibered confocal fluorescence microscopy images

    2014, Artificial Intelligence in Medicine
    Citation Excerpt :

    These features are frequently related to texture of the images [17–19]. They thus may be described by means of the local binary patterns (LBP)[20] which are known to be competitive texture feature extractor [21–24]. Other classical features can also be generated from gray level co-occurrence matrices (GLCM) [25], and they often achieve state-of-the-art recognition performance [14,26,27].

View all citing articles on Scopus
View full text