Hybrid CNN Based Computer-Aided Diagnosis System for Choroidal Neovascularization, Diabetic Macular Edema, Drusen Disease Detection from OCT Images

Optical coherence tomography (OCT) is a medical imaging method that displays eye biological tissue layers at the micron level. Since OCT image interpretation is time-consuming, different Computer-Aided Diagnosis systems have been developed in the current years for semi / full automatic analysis of OCT images.

In the literature, there are CAD systems that use OCT in the diagnosis and classification of retinal diseases and glaucoma diseases and assist in the decision-making process of the doctor [1-3]. Zhang et al. [1] analyzed multiple ophthalmic diseases such as glaucoma, AMD, DMA using a CAD system. Rasti et al. [2] proposed a multi-scale convolutional mixture of expert (MCME) based CAD system to identify normal retina. With the CAD system developed to distinguish AMD, and DME and normal retinal images, they achieved 98.86% sensitivity and AUC 0.99985. Srinivasan et al. [3] presented a fully automated SVM-based method for the detection of retinal diseases in OCT images. SVM entries are multiscale histograms of gradient identifiers directed as feature vectors. The spectral domain OCT datasets used for cross-validation consist of volumetric scans obtained from 45 subjects. In this data set, there are 15 patients with AMD, 15 patients with DME and 15 normal images. The classifier correctly distinguished 100% of cases with AMD, 100% with DME and 86.67% of normal subjects. Lee et al. [4] showed whether deep learning can be used to distinguish normal OCT images from images from patients with AMD. Developed to classify images as normal and AMD, the developed CAD system is a deep neural network based on VGG16. At the image level, ROC and accuracy are 87.63% and 92.78%, respectively. At the macular level, OCTs are classified with 88.98% ROC and 93.83% accuracy, respectively. Liu et al. [5] proposed a machine-learning based method to identify normal macular and three types of macular pathology. The proposed system recognizes the presence of macular hole, macular edema, and AMD in the fovea-centered OCT segment. AUC value of 326 OCT images classified by SVM was obtained 0.93. Lemaître et al. [6] examined the problem of automatically classifying Spectral Domain OCT data for automatic identification of patients with DME compared to normal people. The proposed classification consists of five different steps. The method is a combination of Local Binary Patterns (LBP) features and various mapping strategies, along with various preprocessing steps. The developed method was tested on cohort of 32 patients using linear and nonlinear classifiers. In the experimental results, normal and DME patients were classified with 81.2% sensitivity and 93.7% specificity, respectively. Das et al. [7] presented a new method for drusen (early stage) and CNV (late-stage) from DME and AMD from healthy OCT images. The method includes a multi-scale deep feature fusion-based classification approach. The MDFF fetches the inter-scale variations in images for introducing discriminative and complementary information to the classifier. The method was analyzed using 84,484 OCT images. As a result of the classification, an average of 99.6%, 99.87%, and 99.6% sensitivity, specificity, and accuracy were obtained in the test set. Feng et al. [8] proposed an efficient multi-class 3D retina OCT image classification network called VinceptionC3D. VinceptionC3D is developed from basic C3D with the addition of advanced 3D inception modules. Using 873 OCT images, the proposed method was tested with three different approaches. The average accuracy achieved is 89.35%, 92.09%, and 94.04%, respectively. Albarak et al. [9] proposed a decomposition-based approach with local feature extraction to detect the presence of AMD in OCT images. The proposed technique was applied to 140 volumetric OCT images and evaluated with tenfold cross-validation and the AUC value was obtained 94.4%. Venhuizen et al. [10] designed proposed a machine learning system to automatically rate the AMD severity stages. Classified using 3265 OCT data with early, moderate, or advanced AMD symptoms, the level of AMD was analyzed using ROC analysis and Cohen's κ statistics (κ). The results were compared with the results determined by the two experts. AUC 0.98, sensitivity 98.2%, specificity 91.2%, and κ = 0.713 values were obtained. Sun et al. propose a framework for automated detection of dry AMD, and DME from OCT images. In the proposed framework, a general method is provided to automatically align and trim the retinal regions. Then, spherical representations of the images were obtained using sparse coding and a spatial pyramid. Finally, a multi-class linear SVM was preferred for classification. Two data sets were used to demonstrate the accuracy of the algorithm. For the first dataset, DME, AMD and normal subjects were classified with 100%, 100% and 93.33% accuracy, respectively. The second dataset was classified with 99.67%, 99.67% and 100% accuracy for DME, AMD and normal images, respectively Ref. [11]. Qui and Sun [12] offered self-supervised iterative refinement learning to develop a type of macular 3D OCT image classification algorithm. The presented methods were tested on a clinical data set and a general data set with 5-fold cross-validation. The proposed method provides 89.74%, 94.87% and 93.18% specificity, sensitivity, and accuracy, respectively, in the clinical data set. In the public data set, the results of the three metrics are 90.43%, 98.22%, and 95.88%. Perdomo et al. [13] gave the CNN-based OCT-NET architecture. The architecture is a customized CNN for processing scans from OCT volumes. OCT-NET was applied to the classification of the three states in Spectral Domain Optical Coherence Tomography-OCT volumes. In addition, the proposed model includes a feedback phase that highlights the scanning areas to support the interpretation of the results. The model has been tested on general SERI-CUHK and A2A SD-OCT datasets, including healthy, DR, DME and AMD. SERI + CUHK and A2A SD-OCT datasets were used in the analysis of the proposed method, obtaining 93% sensitivity and AUC 0.99. Fang et al. [14] developed an iterative fusion convolutional neural network for the automatic classification of OCT images. It uses an iterative fusion approach that uses features derived from convolutional layers to ensure the accurate classification of OCT images. The developed method showed better results than traditional CNN and well-known OCT classification methods. Pekala et al. detailed the latest applications of deep learning (DL) methods in the automatic fine-grained segmentation of the spectral field OCT images of the retina. Also, they have also described a new method that combines purely convolutional networks (FCN) with the Gauss Process after processing. In the proposed approach, performance comparisons between other machine learning, such as human clinicians and graphic-based approaches, were made. The approach was tested in an OCT dataset of mild non-proliferative diabetic retinopathy [15]. Acharya et al. [16] proposed a discrete wavelet transform (DWT), Radon transform (RT) and discrete cosine transform (DCT) based system for the detection of DME. In the system, sinograms were obtained from fundus images and subjected to RT. Then, DWT was applied to the sinograms to extract the wavelet coefficients. Finally, DCT was applied to obtain the 2D-DCT coefficients. In addition, coefficients are arranged in zig-zag style and converted into a 1D vector. This 1D signal was subjected to region-sensitive discriminant analysis. Finally, various controlled classifiers have been used to classify three classes using important features. The technique has provided 100% and 97.01% classification accuracy for MESSIDOR databases, using two and seven key attributes, respectively. Dufour et al. [17] proposed applying a random forest to detect the drusen positions in the retinal pigment epithelium. While the proposed method kept our computational complexity to a minimum, 0.94 AUC was achieved in the test data set. Prasath and Ramya [18] proposed a robust and adaptable algorithm for drusen detection based on Gray-Level Co-occurrence Matrix-based textural properties, reducing scan time. Localization of the Optical Disc (OD) and blood vessels using morphological operators helps to increase the accuracy of the classifier. The proposed algorithm determines the macular region based on the location of the vascular passages and OD. The proposed system provides segmentation accuracy above the ROC curve of 98.05%. Yang et al. [19] reported an automated quantification algorithm design for choroidal neovascularization (CNV) in the context of neovascular AMD based on OCT Angiography images. The algorithm is based on segmentation and development methods, including Frangi filter, Gabor wavelets and Fuzzy classification. Puspitawati and Chamidah [20] classified the choroidal neovascularization using a statistical modeling approach based on a local linear estimator. According to the classification results, the accuracy is 90%.

1.1 Motivation

In this paper, a doctorate assists CAD system for classification of Normal, Choroidal Neovascularization, DME, and Drusen OCT images is proposed. The method used is based on deep learning algorithm to categorize OCT images into CNV, Drusen, DME, and normal categories. The proposed system doesn’t require any additional feature extraction or noise filtering on OCT images. SVM is used instead of Softmax layer in recommended CNN models. OCT images are given as input to CNN architectures and the properties in the last pooling layer of CNN are classified with SVM.

1.2 Organization of paper

The rest of the paper is organized as follows to classify OCT images with the CNN-SVM model. The data and features used in Section 2 are presented. The CAD system and Alexnet-SVM structure proposed in Section 3 are detailed and the results are given. The results obtained in Section 4 are presented. In Section 5, literature comparisons are given and discussed. Finally, a brief evaluation is given in the Conclusion section.

Computer-aided diagnosis is an interdisciplinary technology that helps doctors interpret medical images, combining artificial intelligence and computer vision. Comprehensive analysis and evaluation of images such as MRI and ultrasound in a short time is important for the expert. CAD systems process digital images to highlight prominent parts such as images and possible diseases, and support the decisions taken by the doctor. The CAD system proposed in this paper is a system that helps distinguish DME, CNV, Drusen and Normal OCT images. The formation of new blood vessels in the back of the eye and leakage in these vessels is called choroidal neovascularization. DME is also a form of vision loss for diabetic people. Extracellular substance deposits under the retinal pigment epithelium are defined as drusen. Drusen is usually studied in the context of its relationship to AMD. Some varieties of drusen are considered a normal result of aging, while others can be seen at a younger age. Also, some types of drusen are associated with different hereditary and systemic diseases. It is very important to treat such diseases in a timely manner and to prevent visual impairment and to detect early stages of CNV, DME and Drusen. For this purpose, the CAD system in Figure 2, which can classify OCT images, is proposed in the paper.

In the training phase shown in Figure 2.a, OCT images are given as input to the Alexnet-SVM model and classification accuracy is determined. In Figure 2.b the test structure of the system classifying DME, CNV, Drusen and normal OCT images can be seen. The result obtained by the system is decided by doctor control. The Alexnet-SVM architecture used in the proposed CAD system is as in Figure 3. In this proposed model, 6x6x256 = 9216 features in the last pooling layer, where the properties of OCT images are obtained, are classified by SVM. This proposed model is a general model and other CNN models can be used instead of Alexnet.

Krizhevsky et al. [23] AlexNet, which is a convolutional neural network developed by, consists of 8 layers. The image of 227x227x3 size given to AlexNet's input is reduced to 55x55x96 by taking the convolution window 11 × 11 and Stride value 4. This layer is followed by the max pooling layer. The data of 27x27x96 size obtained in this layer is reduced to 6x6x25 by applying 5x5 convolution, Max pooling and 3 times 3 x 3 convolution respectively. The last layer is the max pooling layer. This layer contains the properties of the OCT images given to the input. Brief definitions of layers and SVM classification algorithm used in Alexnet architecture are as follows.

2-1.png

2-2.png

Figure 2. The proposed CAD system

3.png

Figure 3. Architecture of Alexnet-SVM

Convolution: This layer detects the features of the image. It applies some filters to the image to extract low and high level features in the image. Output size is calculated as in Eq. (1) with the help of input volume size (W), the receptive field size of the convolutional layer neurons (F), applied stride (S) and the amount of zero padding used (P) on the border.

$\text { Output }_{\text {size }}=\frac{W-F+2 P}{S}+1$     (1)

RELU: The rectified linear unit (RELU) is a nonlinear function. While the ReLU function, which is faster than functions such as Tanh, takes the value 0 for negative inputs, x takes the value x for positive inputs. It is calculated as in Eq. (2).

$f(x)= \begin{cases}0 & x<0 \\ x & x \geq 0\end{cases}$     (2)

Max Pooling: The pooling layer is usually used after a convolutional layer. The pooling layer is responsible for reducing the size of features. It aims to extract dominant features from images to speed up computation. In addition to maximum pooling, it is also used for average pooling and L2-norm pooling. In practice, in the variation commonly used in the maximum pooling layer, F = 3, S = 2 is used. This is also called overlapping pooling. It is calculated as in Eq. (3).

$\text { Output }_{\text {size }}=\frac{W-F}{S}+1$     (3)

Normalization: Normalization layers are used between the convolution layer and the nonlinear layer to speed up network training and reduce network start sensitivity. Local Response Normalization (LRN) is used in AlexNet to help with generalization. Eq. (4) LRN gives the calculation of normalization.

$b_{x, y}^{i}=a_{x, y}^{i} /\left(k+\alpha \sum_{j=\max \left(0, i-\frac{n}{2}\right)}^{\min \left(N-1, i+\frac{n}{2}\right)}\left(a_{x, y}^{j}\right)^{2}\right)^{\beta}$     (4)

where, α: activation of a neuron, i: ith kernel, N: total number of kernels, α, β, k, n: are hyper parameters.

Support Vector Machine: SVM is one of the popular supervised learning methods used in classification problems. However, it is mostly used in classification problems. The main goal is to find the most suitable plane or hyperplane so that it can distinguish into two classes.