Multi-parametric optic disc segmentation using superpixel based feature classification☆
Introduction
Nowadays, people are unaware of the visual impairment and blindness lesions of an eye that are macular degeneration, hypertension, glaucoma, and diabetic retinopathy (Lee, Wong, Sabanayagam, 2015, Soomro, Khan, Khan, Gao, Paul, Zheng, 2018). Glaucoma is an incurable eye pathology around the optic disc (OD) that badly damages the optic nerves and leads to vision loss. Once started, it cannot be cured but can be prevented by diagnosing at its early stage (Soomro, Gao, Khan, Hani, Khan, Paul, 2017, Soomro, Khan, Gao, Khan, Paul, 2017). According to an estimate, it becomes the most crucial and leading cause of vision blindness, it may affect about 80 million people by 2020 (Khan, Khan, Bailey, Soomro, 2018, Khan, Khan, Soomro, Mir, Gao, 2017, Quigley, Broman, 2006). The progression rate of this disease is slow but occurs gradually over an extended time interval. The symptoms of glaucoma appear after an extensive time period when the disease is quite advanced. It is incurable but the growth rate of its progression rate have been regressed by proper treatment if it is diagnosed at its primary stage. The glaucoma progression is symptomized from the few signs of vision loss. In Australia, 50% of peoples are unfamiliar with glaucoma (Maldhure & Dixit, 2015). The OD is defined by its central bright region called optic cup (OC) and the corresponding outer peripheral region called the neuroretinal rim, as shown in Fig. 1. The cup to disc ratio (CDR), which is the ratio of vertical cup diameter (VCD) to vertical disc diameter (VDD) is considered as a vital factor in glaucoma screening (Almazroa, Burman, Raahemifar, & Lakshminarayanan, 2015), due to its directly proportional relationship with glaucoma occurrence. The segmentation of the OC has been demonstrated to rely on accurate localization of the OD in literature (Yin et al., 2012). On similar lines, in this work, we argue that the correct localization of the OD is vital for eye pathology screening and diagnosis.
Expert and intelligent systems continue to contribute towards scientific and technological advances in numerous fields from medical applications, telecommunications, economics, transportation, and surveillance. An important milestone for intelligent machine vision systems is to match or surpass human vision performance. Expert automated vision systems have been successfully applied to disease diagnosis: recent works include real-time detection of tuberculosis using an intelligent mobile-enabled expert system (Shabut et al., 2018), deep neural network based recommender system for skin lesion extraction (Soudani & Barhoumi, 2019) and a convolutional neural network based segmentation framework for breast tumor classification (Rouhi, Jafari, Kasaei, & Keshavarzian, 2015).
The multifaceted challenges in OD localization include deformable shape, variation in color, OD boundary discontinuities due to blood vessels and glaucoma pathology such as peripapillary atrophy (PPA) (Jonas, Fernández, & Naumann, 1992), and ISNT rule (Harizman et al., 2006). Most of the unsupervised and supervised methods for OD boundary detection are gradient dependent approaches. Hence, they struggle to capture true OD boundary in cases where peripapillary atrophy (PPA) makes the OD boundary discontinuous or the vascular structure originating from the OD could misguide the segmentation. Also, the supervised approaches for OD localization (Fan et al., 2018) typically give less attention to important stages of the learning pipeline, including preprocessing, appropriate feature selection and data imbalance problem in medical images.
To this end, we proposed a supervised approach for OD localization which is independent of boundary information, the deformable shape of the OD and the irregularities in its size. To enhance the OD region and suppress the vessel background, appropriate image preprocessing is applied to increase the generalization ability of the classifier. The proposed method utilizes the statistical and textural properties of regions to learn a discriminative classifier from positive and negative regional samples. Pixel-accurate ground truth annotations are employed to generate training instances. In this work, various discriminative classifiers are trained and the Random Forest classifier is selected as the proposed approach due to its better generalization performance. The proposed method can robustly localize OD with high accuracy, which makes it a suitable candidate for glaucoma screening and diagnosis. The specific contributions of this work are:
- 1.
The problem of OD segmentation is modeled as a regional classification problem through a systematic classification pipeline.
- 2.
Image regions are characterized by statistical as well as textural properties in a well-formulated classification framework to make it robust against multifaceted challenges in OD localization.
- 3.
A thorough comparison of regional classifiers is presented for OD localization in retinal fundus images. Our results demonstrate the improved generalization of the Random Forest classifier over other classification based methods and unsupervised counterparts.
The rest of this paper is organized as follows. A detailed review of the relevant methods is presented in Section 2. The proposed method’s main steps for OD localization are presented in Sections 3–8 that comprise of the preprocessing, superpixel segmentation, feature extraction and selection and the classification of each superpixel by using machine learning tools. Sections 9 and 10 comprise of experimental setup that contains datasets description, parameter selection, and comparative experimental results. The discussion and conclusions are described in Sections 11 and 12, respectively.
Section snippets
Literature review
Computerized OD localization and segmentation is still considered as a highly challenging task and an open problem in medical image analysis and diagnosis. The obvious challenges include the ever-varying physical characteristics of OD (such as size, structure, vessel structure, color) and various glaucoma pathology, as shown in Fig. 2. Considering only the independent scanning protocols, all of the mentioned properties are unknown. Many OD and eye lesion detection algorithms have been developed
Method
The proposed method uses DRIONS and also further evaluated on MESSIDOR and ONHSD dataset. The main block diagram of the proposed work is depicted in Fig. 3, the details of these block are described in the following subsections.
Feature extraction
Feature extraction and normalization steps are very important to make the classification based approach more robust for OD segmentation that are detailed below.
Feature normalization
For classification-based approaches, feature normalization (Cao, Stojkovic, & Obradovic, 2016) is important because most of the classifier work on distance based schemes. It is necessary to obtain a robust classification by normalizing the feature variable in a specific range. There are many approaches that are used to standardize the features by mean and histogram normalization. In this paper, mean normalization is used to standardize the features. Given feature f, the mathematical formulation
Feature selection
Feature selection is an essential step for supervised methods. Feature selection helps in fighting with the curse of dimensionality. Good features improve the prediction performance of the classifier. In this paper, features are selected on the basis of features mutual information (MI(a, b)) to find the minimum redundancy between feature sets (Peng, Long, & Ding, 2005). Formally, the MI is defined as follows:where ℵ(a, b) is the joint mass probability of
Classification
In supervised classification the model is constructed by learning from data along with its annotated labels. The learned model is then evaluated on unseen data. In the proposed work, we employ four supervised classification methods for optic disc segmentation and present a comprehensive comparison on benchmark datasets.
The Support Vector Machine (SVM) classifier is a supervised classifier used for classification (Boser, Guyon, & Vapnik, 1992). It can be used for both linear and nonlinear
Ellipse fitting
The image segmented by the classifier is binarized for further processing. As vessels are also present in the OD region, they also affect the OD region during binarization. In addition, the boundaries of OD can be affected by the nonuniform illumination. To mitigate these effects and estimate the true OD region, the largest connected object is obtained and its boundary is used as the raw estimation. The best-fitted ellipse is computed as the OD boundary (Zhang et al., 2009). The ellipse fitting
Experimental analysis
In this section, the results of four nonlinear classifiers are evaluated on the publicly available retinal fundus datasets DRIONS, ONHSD and MESSIDOR to test the robustness of the proposed method. This section reports our experiments including parameter selection and results evaluation.
Hierarchical classification
In this section we present the quantitative comparison of the classifiers for the task of OD segmentation in terms of the selected measures on all datasets followed by visual results of the proposed method on representative images from benchmark databases. Next, a comparison of the proposed method with the state-of-the-art methods is presented.
Discussion
This paper presents a supervised approach for OD localization. The presented approach used DRIONS database, ONHSD and MESSIDOR data for evaluation and relies on intensity, texton-map histogram and fractal features.
In the progression of feature computation process, we considered intensity, texton-map, and fractal features. Important parameters required at the feature computation stage were sought empirically as optimization of these parameters is beyond the scope of this work. Future work will
Conclusion
This paper presented a regional classification framework for accurate localization and segmentation of the optic disc. The modeling of the optic disc segmentation as a region-based classification by utilizing multi-modality attributes proved to be robust against the multifaceted challenges of optic disc detection. The results demonstrate that the proposed method is resilient against the highly varying nature of optic disc appearance as compared with other expert and intelligent systems, which
Acknowledgments
The authors would like to specially thank the management teams of the MESSIDOR, DRIONS and ONHSD databases for providing retinal fundus image databases along with the manual expert annotations. These database were download from there publicly available links. The authors would like to thank Higher Education Commission Pakistan for funding the research reported in this paper under grant number 21-2020/SRGP/R&D/HEC/2018.
References (46)
A computational approach to edge detection
Readings in computer vision
(1987)- et al.
A decision-theoretic generalization of on-line learning and an application to boosting
Journal of Computer and System Sciences
(1997) - et al.
An intelligent mobile-enabled expert system for tuberculosis disease diagnosis in real time
Expert Systems with Applications
(2018) - et al.
Computerised approaches for the detection of diabetic retinopathy using retinal fundus images: A survey
Pattern Analysis and Applications
(2017) - et al.
Bias in random forest variable importance measures: Illustrations, sources and a solution
BMC Bioinformatics
(2007) - et al.
Model-based optic nerve head segmentation on retinal fundus images
Engineering in medicine and biology society, EMBC, 2011 annual international conference of the IEEE
(2011) - et al.
Localization and segmentation of optic disc in retinal images using circular hough transform and grow-cut algorithm
PeerJ
(2016) - et al.
Frequency-tuned salient region detection
2009 IEEE conference on computer vision and pattern recognition
(2009) - et al.
Slic superpixels compared to state-of-the-art superpixel methods
IEEE Transactions on Pattern Analysis and Machine Intelligence
(2012) - et al.
Optic disc and optic cup segmentation methodologies for glaucoma image detection: A survey
Journal of Ophthalmology
(2015)
Deep convolutional neural networks for brain image analysis on magnetic resonance imaging: a review
Artificial Intelligence in Medicine
A training algorithm for optimal margin classifiers
Proceedings of the fifth annual workshop on computational learning theory
Random forests
Machine Learning
A robust data scaling algorithm to improve classification accuracies in biomedical data
BMC Bioinformatics
Automatic optic disc segmentation with peripapillary atrophy elimination
Engineering in medicine and biology society, EMBC, 2011 annual international conference of the IEEE
Convolutional neural network and texture descriptor-based automatic detection and diagnosis of glaucoma
Expert Systems with Applications
Optic disk detection in fundus image based on structured learning
IEEE Journal of Biomedical and Health Informatics
Joint Optic Disc and Cup Segmentation Based on Multi-Label Deep Network and Polar Transformation
IEEE Transactions on Medical Imaging
Digital image processing
The ISNT rule and differentiation of normal from glaucomatous eyes
Archives of Ophthalmology
3d surface tracking and approximation using Gabor filters
Fundamentals of digital image processing
Glaucomatous parapapillary atrophy: Occurrence and correlations
Archives of Ophthalmology
Cited by (56)
C2FTFNet: Coarse-to-fine transformer network for joint optic disc and cup segmentation
2023, Computers in Biology and MedicineRecent trends and advances in fundus image analysis: A review
2022, Computers in Biology and MedicineOptic disc detection and segmentation using saliency mask in retinal fundus images
2022, Computers in Biology and MedicineCitation Excerpt :On the other hand, Pathan, et al. [19] outperformed our method in terms of sensitivity, using the decision tree classifier which was tested on only one public database (Rim One) with low specificity and accuracy. The proposed system consistently achieves comparable or better results in terms of sensitivity when compared to the literature methods that tested the Drions database, except for Rehman et al. [18] and Khan et al. [35] who obtained lower specificity, accuracy, and Dice coefficient values. In addition, Khan et al. [35] got lower accuracy (99.60%) and sensitivity (96.49%) for the Drishti database by comparing with our method which has an accuracy of 99.62% and a sensitivity of 98.20%.
Automatic detection and segmentation of optic disc using a modified convolution network
2022, Biomedical Signal Processing and Control
- ☆
An Explicit Method for Localization of Optic disc from Retinal fundus Images.