An anomaly detection approach for the identification of DME patients using spectral domain optical coherence tomography images

Highlights

• We propose a method to detect abnormal B-scans in SD-OCT volumes.

• We develop an anomaly detection approach for classification of SD-OCT volumes.

• The proposed method achieves high sensitivity and specificity compared to other supervised classification methods.

Abstract

This paper proposes a method for automatic classification of spectral domain OCT data for the identification of patients with retinal diseases such as Diabetic Macular Edema (DME). We address this issue as an anomaly detection problem and propose a method that not only allows the classification of the OCT volume, but also allows the identification of the individual diseased B-scans inside the volume. Our approach is based on modeling the appearance of normal OCT images with a Gaussian Mixture Model (GMM) and detecting abnormal OCT images as outliers. The classification of an OCT volume is based on the number of detected outliers. Experimental results with two different datasets show that the proposed method achieves a sensitivity and a specificity of 80% and 93% on the first dataset, and 100% and 80% on the second one. Moreover, the experiments show that the proposed method achieves better classification performance than other recently published works.

Introduction

Diabetic retinopathy (DR) and diabetic macular edema (DME) are leading causes of vision loss worldwide [1]. As the number of people affected by diabetes is expected to grow exponentially in the next few years [2], [3], early detection and treatment of these retinal diseases is becoming a major health issue in developed countries. Currently, the two main tools used for the screening of retinal diseases are fundus photography and optical coherence tomography (OCT) devices. The former type of device allows capturing the inner surface of the eye, including retina, optic disc, macula, and other retinal tissues using low power microscopes equipped with an embedded camera system [4]. The latter is based on infrared optical reflectivity and produces cross-sectional and three-dimensional images of the central retina. The new generation of OCT imaging, namely Spectral Domain OCT (SD-OCT) offers higher resolution and faster image acquisition over conventional time domain OCT [5]. SD-OCT can produce from 27,000 to 40,000 A-scans/second with an axial resolution ranging from 3.5 µm to 6 µm [6].

While many previous works in literature are dedicated to the analysis of fundus images [7], [8], in this work we focus on the automatic analysis and classification of SD-OCT images. Indeed, the use of fundus photography is limited to the detection of signs which are visible in the retinal surface, such as hard and soft exudates, or micro-aneurysms, and fundus photography cannot always identify the clinical signs such as cysts not visible in the retinal surface. Moreover, contrary to fundus photography, SD-OCT provides quantitative measurements of retinal thickness and information about cross-sectional retinal morphology.

The main, common, feature associated with DME is an increase of the macular thickness in the central area of the retina [9]. Therefore, several methods have been proposed which are based on the segmentation of retinal layers and the identification of signs of the disease such as intraretinal and subretinal fluid [10], [11]. For example, Quellec et al. [10] proposed a method that starts with the segmentation of 10 retinal layers and the extraction of texture and thickness properties in each layer. Then, local retinal abnormalities are detected by classifying the differences between the properties of normal retinas and the diseased ones. Chen et al. [11] proposed a method for 3D segmentation of fluid regions in OCT data using a graph-based approach. The method is based on the segmentation of retinal layers and the identification of potential fluid regions inside the layers. Then, a graph-cut method is applied to get the final segmentation using the probabilities of an initial segmentation based on texture classification as constraints. Although methods based on a prior segmentation of retinal layers have reported good detection accuracy, this first step is still difficult and subject to errors [12], [13]. Moreover, as mentioned by Lee et al., retinal thickness measurement differences produced by different algorithms are important and it is not always possible to compare retinal thicknesses among eyes in which thickness measurements have been obtained by different systems [14].

In this paper, we focus on methods based on direct classification of SD-OCT data without prior segmentation of retinal layers. Several methods have recently been proposed in literature for the classification of SD-OCT data and the identification of patients with retinal diseases versus normal patients. Liu et al. [15] proposed a method based on local binary patterns (LBP) and gradient information for macular pathology detection in OCT images. The method uses a multi-resolution approach and builds a 3-level multi-scale spatial pyramid of the foveal B-scan for each patient. LBP and gradients are then computed from each block at every level of the pyramid, and represented as histograms. The obtained histograms are concatenated into a global descriptor whose dimension is reduced using principal component analysis (PCA). Finally a support vector machines (SVM) classifier is used for classification. With a dataset of 326 OCT volumes, the method achieved good results in detecting OCT scans containing different pathologies such as DME or age-related macular degeneration (AMD), with an area under the ROC curve (AUC) value of 0.93.

Another method using gradient information and SVM is proposed by Srinivasan et al. [16] to distinguish between DME, AMD, and normal SD-OCT volumes. The method is based on pre-processing to reduce the speckle noise in OCT images and flattening of the images to reduce the variation of retinal curvature among patients. Histograms of oriented gradients (HOG) features are then extracted from each B-scan of an OCT volume and a linear SVM is used for classification. Note that the method classifies each individual B-scan into one of the three categories, i.e. DME, AMD and normal, and classifies an OCT volume based on the number of B-scans in each category. On a dataset of 45 patients containing 15 normal subjects, 15 DME patients and 15 AMD patients, the method achieved a correct classification of 100%, 100% and 86.67% for AMD, DME and normal cases, respectively. Note that no specificity or sensitivity values are reported in this study.

Anantrasirichai et al. [17] proposed a method for OCT images classification using texture features such as LBP, grey-level co-occurrence matrices and wavelet features, together with retinal layer thickness. The method relies on SVM to classify retinal OCT images of normal patients against patients with glaucoma. A classification accuracy of about 85% is achieved with a dataset of 24 OCT volumes. Albarrak et al. [18] first decomposed the OCT volume into sub-volumes, and extracted both LBP and HOG features in each sub-volume. The features from the sub-volumes are concatenated into a single feature vector per OCT volume, and PCA is applied for dimensionality reduction. Finally, a Bayesian network classifier is used and the method is tested with 140 OCT volumes of normal and AMD patients. The method achieved an AUC value of 94.4%.

The authors in Ref. [19] proposed a method based on the bag-of-words (BoW) approach for OCT images classification. The method starts with the detection and selection of few keypoints in each individual B-scan. This is achieved by keeping the most salient points corresponding to the top 3% of the vertical gradient values. Then, an image patch 9 × 9 pixels in size is extracted around each keypoint, and PCA is applied to transform every patch into a feature vector of dimension 9. These feature vectors are then used to create a visual vocabulary using k-means clustering algorithm, which is used to represent each OCT volume as a histogram of the word's occurrences. Finally, this histogram is used as feature vector to train a random forest classifier. On the task of classifying OCT volumes between AMD and normal cases, the method achieved an AUC of 98.4% with a dataset of 384 OCT volumes. Another approach based on BoW model is proposed by Lemaître et al. [20] for automatic classification of OCT volumes. The method is based on LBP features to describe the texture of OCT images and dictionary learning using the BoW approach. In this work, the authors extracted both 2D and 3D LBP features to describe OCT volumes, and show that 3D LBP features perform better than 2D LBP features from each B-scan. The features are used to create a visual vocabulary using the BoW approach, and a random forest classier is employed for classification. The method achieved a specificity and a sensitivity of 75% and 87.5% respectively, with a dataset of 32 OCT volumes.

In this paper, we propose a novel method for automatic identification of patients with retinal diseases, such as DME, versus normal subjects. Our approach is based on modeling the appearance of normal OCT images with a Gaussian Mixture Model (GMM) and detecting abnormal OCT images as outliers. Then, the classification of an OCT volume as normal or abnormal is based on the number of detected outliers in the volume. The main contributions of our paper are as follows:

• We propose a method to model the global appearance of normal OCT volumes using a Gaussian mixture model (GMM).

• We use an anomaly detection approach to identify abnormal B-scans as outliers to the GMM model, and finally detect unhealthy OCT volumes based on the number of abnormal B-scans.

Our approach differs from previous works in that our abnormal B-scan detection method does not require a training set with manually identified normal and abnormal B-scans which is a tedious and time consuming task.

The rest of this paper is organized as follows. In Section 2, we describe the proposed anomaly detection approach in details. Experiments and results are discussed in Section 3. Finally, concluding remarks are drawn in Section 4.