Scattering features for lung cancer detection in fibered confocal fluorescence microscopy images
Introduction
Lung cancer is one of the most widespread forms of cancer and a major cause of premature death. Survival of lung cancer is highly associated with early diagnosis [1]. It follows that it is of primary importance to develop methods to help to diagnose cancer. Numerous imaging methods have been developed in order to improve the detection of these early cancerization stages [2]. Medical imaging tools for lung cancer diagnosis are mostly based on inspections of the lung and the chest [3] using X-ray-based systems including computed tomography [4], [5]. For centrally located lung cancer, bronchoscopy is an essential tool for diagnosis. Until recently, the early diagnosis of lung cancer relied primarily on a rough inspection of the bronchial mucosae during an endoscopic procedure and on pathology examination of biopsy samples. Classic fiberoptic bronchoscopy, using white-light illumination, has repeatedly shown a low sensitivity for the detection of the early – presumably curable – lesions such as carcinoma in situ [6]. In particular, some techniques have emerged over the past five years that extend the field of exploration of bronchoscopy to the distal lung and to the cellular level, among which fibered confocal fluorescence microscopy (FCFM) [7], [8].
The FCFM technique (also referred to as probe-based confocal laser endomicroscopy, especially in gastrointestinal imaging) 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, whose applications for lung exploration are currently under investigation, can replace lung biopsy in the future and may prove to be helpful in the diagnosis of a number of diseases [7]. FCFM has other challenging applications, such as the imaging of colonic polyps in gastroenterology, for which image retrieval and classification works are ongoing [9]. FCFM imaging is thus able to provide in vivo cellular images of the lung; for this reason, it opens the road to novel methods for early diagnosis of lung diseases including cancer. This is one of the objectives of this paper: to assess the feasibility of computer-aided lung cancer detection based on FCFM images.
Computer-aided diagnosis systems for lung diseases or cancer based on imaging usually involve image processing techniques and machine learning approaches. Many approaches focus on nodule detection in chest radiographs [10], [11] or in computed tomography (CT) [12], [13], [14]. CT is often coupled with advanced numerical techniques for extracting relevant features from the images as well as with state-of-the-art classifiers [15], [16]. Usually, the systems are based upon a two-stage approach: feature extraction and classifier training [11], [17]. The first stage aims at building relevant features that help in discriminating tissues while in the second stage, the classifier learns from examples to automatically assign a class, typically normal or abnormal, from a set of features extracted from an image. These features are frequently related to texture of the images [17], [18], [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], [22], [23], [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]. However, we depart from this mainstream use of LBP for texture features and instead investigate the benefit of using a novel wavelet-based transform, named scattering transform, for building discriminative texture features.
Scattering transform can be understood as successive applications of a wavelet-based linear filtering followed by a modulus operator and a local averaging. This cascade of wavelet transform and non-linearity operators makes the scattering transform locally invariant to translation and stable to small deformations (shear, rotations or any other displacement field) [28], [29], [30]. Here, stability to deformations means that the norm of the difference between the scattering transform of an image and its deformed version is bounded by a term which depends on the norm of the image and a constant depending on the deformation. Therefore, this theoretical property guarantees that the scattering representation of an image does not change too much if the image is deformed, which is a positive property for reducing intra-class variability in classification problems. As such, features resulting from this transform are expected to be more effective than a classical wavelet transform. The other objective of this work is to assess the relevance of scattering features for lung cancer detection in images. As the FCFM dataset we use is novel, we have also evaluated the value of these features on other well-known, publicly available medical image datasets (the 2D-Hela and the Pap smear datasets).
As far as we know, this work presents the first application of scattering transform to medical images. The contributions we present are three-fold:
- •
First, we assess the feasibility of lung cancer diagnosis from FCFM images. FCFM imaging is a rather recent imaging technique and its importance and applicability are yet to be explored. In this paper, we show that this image modality can help detecting lung cancer.
- •
Second, we show that scattering coefficients are highly discriminative features for this FCFM-based lung cancer classification problem. When compared to one of the state-of-the-art features, the LBP feature, and its recent variant local quinary pattern (LQP) [24], they achieve strongly competitive results, yielding the best discrimination performance. Comparisons with GLCM-based features also show that scattering features are efficient.
- •
Finally, in order to achieve insight on the discriminative power of scattering representation, we further analyze its behavior on two other medical image classification problems. The findings suggest that i) the choice of the wavelet has an effective impact on the classification performance, ii) a generative classifier adapted to scattering features does not perform necessarily better than a standard discriminative classifier like a support vector machine (SVM), iii) coupling LBP or LQP and scattering features by concatenation may further enhance classification performances, and iv) fusing classifiers trained with single LBP, LQP and scattering features yield to slightly better performances than a single classifier trained with concatenated features.
The paper is organized as follows. Section 2 briefly presents the FCFM imaging technique as well as the image dataset used in the experiments. A discussion on the scattering transform is given in Section 3. This section also briefly recalls how LBP and LQP features are constructed and shortly introduces the classifier that is used for learning a decision function. Experimental analyses are reported in Section 4.
Section snippets
Materials
FCFM bronchi images are acquired by the application of a probe tip onto the bronchial wall. Even if cancer significantly alters the bronchial epithelial layer, the absence of epithelial cell visualization does not allow FCFM to differentiate between the different grades of progression of the pre-cancer bronchial lesions such as metaplasia, dysplasia and carcinoma in situ. To be successfully applied to the exploration of pre-cancer/cancer bronchial epithelial layer, the FCFM technique needs to
Scattering representation of images
Scattering transform has recently been introduced by Bruna et al. [28] in order to build representations of images and signals that are stable to deformations. Informally, a scattering transform recursively applies a cascade of wavelet decomposition and modulus operator and in some sense, it mimics convolutional neural networks as it cascades several layers of filtering [31]. We will now provide a more formal definition of the scattering representation of an image. More details can be found in
Experimental results and discussions
The aims of the experimental results are two-fold. Primarily, we want to empirically evaluate the efficiency of scattering-operator-based features on classical and well-known medical imaging classification problems. Afterwards, we focus on the FCFM dataset and provide an in-depth empirical analysis of scattering features for this lung imaging cancer detection problem. The Matlab code used for producing these results is available on the author's website.1
Conclusions
The objective was to assess the feasibility of lung cancer diagnosis with FCFM imaging techniques. We have shown that when features obtained from scattering transform are extracted from these images, it is possible to learn a classifier able to achieve image recognition rate score as high as 80% in a Leave-One-Patient-Out setting. Similar performances, although slightly less successful, can also be obtained using classical feature extraction techniques such as Local Binary Patterns.
Acknowledgment
This work was partly supported by grants from ASAP ANR-09-EMER-001.
References (38)
Imaging and cancer: a review
Mol Oncol
(2008)- et al.
Comprehensive diagnostic bronchoscopy of central type early stage lung cancer
Lung Cancer
(2007) - et al.
A smart atlas for endomicroscopy using automated video retrieval
Med Image Anal
(2011) - et al.
Computer-aided diagnosis of pulmonary nodules using a two-step approach for feature selection and classifier ensemble construction
Artif Intel Med
(2010) - et al.
Local binary patterns variants as texture descriptors for medical image analysis
Artif Intel Med
(2010) - et al.
An SVM-based distal lung image classification using texture descriptors
Comput Med Imaging Graph
(2012) - et al.
Pap smear diagnosis using a hybrid intelligent scheme focusing on genetic algorithm based feature selection and nearest neighbor classification
Comput Biol Med
(2009) - et al.
Early detection of lung cancer: clinical perspectives of recent advances in biology and radiology
Clin Cancer Res
(2001) Imaging lung cancer
Semin Oncol
(1999)- et al.
Computer-aided detection (cad) of lung nodules and small tumours on chest radiographs
Eur J Radiol
(2009)
Computer analysis of computed tomography scans of the lung: a survey
IEEE Trans Med Imaging
Confocal fluorescence endomicroscopy of the human airways
Proc Am Thorac Soc
Bronchoscopic advances: on the way to the cells.
Respiration
Image-processing technique for suppressing ribs in chest radiographs by means of massive training artificial neural network
IEEE Trans Med Imaging
A fully automated method for lung nodule detection from postero-anterior chest radiographs
IEEE Trans Med Imaging
Shape-based computer-aided detection of lung nodules in thoracic ct images
IEEE Trans Biomedical Eng
Patient-specific models for lung nodule detection and surveillance in ct images
IEEE Trans. Med Imaging
Automatic detection and quantification of tree-in-bud (tib) opacities from CT scans
IEEE Trans Biomed Eng
Computer-aided diagnosis of pulmonary nodules on ct scans: improvement of classification performance with nodule surface features.
Med Phys
Cited by (25)
Comparing Probe-Based Confocal Laser Endomicroscopy With Histology. Are We Looking at the Same Picture?
2021, Archivos de BronconeumologiaImage computing for fibre-bundle endomicroscopy: A review
2020, Medical Image AnalysisCitation Excerpt :Commonly used feature descriptors include (i) first order image statistics, (ii) structural information through Skeletonisation, Sobel and Canny Edge Detectors, etc. (iii) Haralick's texture parameters derived through gray Level Co-occurrence Matrices (GLCM), (iv) Local Binary Patterns (LBP) and their variation of Local Quinary Patterns (LQP), and (v) Scale Invariant Feature Transforms (SHIFT). Other less adopted descriptors employed as discriminative features include (i) spatial frequency based features extracted at Fourier domain (Srivastava et al., 2005,2008), (ii) fractal analysis (Ştefănescu et al., 2016), (iii) Scattering transform (Rakotomamonjy et al., 2014; Seth et al., 2016), (iv) Fast Retina Keypoint (FREAK) (Wan et al., 2015), (v) oriented FAST and rotated BRIEF (ORB) (Wan et al., 2015), (vi) Histogram of orientated Gradients (HOG) (Gu et al., 2016; Vo et al., 2017), (vii) textons (Gu et al., 2016), (viii) Local Derivative Patterns (LDP) (Vo et al., 2017), as well as (ix) features extracted from Convolutional Neural Networks (CNN) prior to the fully connected layer employed for computing each class score (Gil et al., 2017; Vo et al., 2017). Leonovych et al. (2018) introduced Sparse Irregular Local Binary Patterns (SILBP), an adaptation of LBPs taking into consideration the sparse, irregular sampling imposed by the imaging fibre bundle on FBEµ images.
Review of recent developments in determining volatile organic compounds in exhaled breath as biomarkers for lung cancer diagnosis
2017, Analytica Chimica ActaCitation Excerpt :1-octene is related to a possible biochemical pathway leading to advanced stages of NSCLC. The metabolic microenvironment of tumor cells can be influenced by cellular oncogenic signaling, metabolic demands of cancer tissue, and changes in the expression of specific enzymes [1,2,4–11,13–17,20–27]. Several studies have found that tumor growth may be accompanied by changes in genes and/or proteins that lead to peroxidation of the cell membrane species and the release of endogenous compounds as VOCs [53–65].
Scalable gastroscopic video summarization via similar-inhibition dictionary selection
2016, Artificial Intelligence in MedicineCitation Excerpt :Many computer-aided endoscopy diagnosis systems have been proposed to assist clinicians in improving the accuracy of medical diagnosis using the images or videos recorded in the inspection of a GT tract. According to the specific lesions, these systems can be classified to handle bleeding [30,31], tumors [32,33], Helicobacter pylori [34], cancer [35,36], Crohn's disease [37] and polyps [38]. Moreover, some other applications include pose detection for endoscopy [39], video segmentation [40] and three-dimensional reconstruction of the digestive wall [41].
Detection and classification of pavement damages using wavelet scattering transform, fractal dimension by box-counting method and machine learning algorithms
2024, Road Materials and Pavement Design