Automatic Detection and Classification of Malarial Retinopathy-Associated Retinal Whitening in Digital Retinal Images

Malarial retinopathy addresses diseases that are characterized by abnormalities in retinal fundus imaging. Macular whitening is one of the distinct signs of cerebral malaria but has hardly been explored as a critical bio-marker. The paper proposes a computerized detection and classification method for malarial retinopathy using retinal whitening as a bio-marker. The paper combines various statistical and color based features to form a sound feature set for accurate detection of retinal whitening. All features are extracted at image level and feature selection is performed to detect most discriminate features. A new method for macula location is also presented. The detected macula location is further used for grading of whitening as macular or peripheral whitening. Support vector machine along with radial basis function is used for classification of normal and malarial retinopathy patients. The evaluation is performed using a locally gathered dataset from malarial patients and it achieves an accuracy of 95% for detection of retinal whitening and 100% accuracy for grading of retinal whitening as macular or non-macular. One of the major contributions of proposed method is grading of retinal whitening into macular or peripheral whitening.

Various medical decision support systems [2,3] have been developed that assist the specialists in diagnosis, prognosis and treatment of various ailments. The fundus images of normal subjects and subjects with disease are used to develop a machine learning system that assists in the decision making process. In case of retinal region, the region is manifested by abnormalities like exudates, MA (Microaneurysms), hemorrhages, colored spots, discoloration and whitening. Although the features extracted for all these abnormalities are different, however, the underlying image processing techniques have similarities that can be exploited in detection of these abnormalities.
Automatic detection and classification of various diseases through analysis of fundus images has interested researchers for many years [4]. Researchers have used the knowledge base developed in various ophthalmological studies for developing computer based statistical models to diagnose various ailments. Most of this research focuses on the development of image processing techniques to detect the abnormalities in the retinal region and then extract discriminatory features from these areas using preprocessing, noise removal, image enhancement and segmentation.
The paper is organized as follows. Section 2 highlights the related work with respect to malarial retinopathy, cerebral malaria and other ophthalmological disorders. Section 3 discusses the proposed methodology of the system. Section 4 highlights the experimental results of the system supported by various analytical and statistical parameters. Finally, Section 5 concludes the paper briefly.  [11]. They were able to obtain accuracies of around 83% using the proposed methods.

RELATED WORK
In fundus images, the localization of optical disk and macula is an important step before further analysis on detection of abnormalities can be done. Similarly, enhancement and segmentation of blood vessels in fundus images is important to localize the areas affected by any retinal abnormalities.
Geometric relationship between the optical disk and the blood vessels was used by Hoover et. al. [12]  to report different signs of MR and also to highlight the significance of retinal whitening [5][6][7][8]. The paper proposes a medical decision support system to automatically detect and classify the retinal whitening and its types using fundus images. The system also distinguishes the normal subjects from the non-healthy subjects. This is the main contribution of this article as it is a unique study carrying an automated system for detection of retinal whitening.

PROPOSED METHODOLOGY
Automatic detection of malarial whitening in fundus images can be vital for early treatment of patients.
Decision support systems can be live-savers in remote locations where specialist care is not available for patients.
In this article, we propose an automatic detection and classification system that detects and classifies the malarial retinopathy associated retinal whitening. The

Pre-Processing
The first step of pre-processing involves basic processing on the raw data before complex processing steps are adopted on the images dataset. Some of the acquired images in the dataset correspond to a different resolution, so in the pre-processing step the images in the dataset are re-sized to a resolution of 1536x2048 for uniformity in all the corresponding steps.

Retinal Whitening Region Localization
Once The removal of these pixels from the image helps in the segregation of retinal whitening pixels in the image. The process for detection of OD is explained in [9] and the same method has been used. High intensity pixels associated with OD are removed from the image using a 80x80 window. After the removal of these pixels, mean based thresholding is performed on the image and multiple features are computed from the image. Fig. 9 shows the colored image and the RGB components of a normal subject after OD removal.  The details of these features is explained in the next section.

Feature Extraction
For an automated system to distinguish between normal subjects and subjects with retinal whitening in fundus

Classification
In where K(X i ,X j ) is called the kernel function. Here y is class label,α is coefficient and b is biased term.    For the detection of blood vessels, 2D Gabor wavelet is used on the inverted green channel of the fundus image. This is followed by multilayer thresholding for vessel segmentation [22]. OD detection algorithm in [18] has already been explained and the same has been used for detection of the pixels corresponding to OD. Minimum distance classifier and Otsu's algorithm [23] are used to segment the OD and the blood vessels from the fundus

Material
To evaluate the performance of the proposed method, a dataset of 22 images has been usedwhich have been ophthalmologist. Fig. 16 shows some of the images from the dataset.

Results
The feature vector extracted from the images dataset contains 48 features. In the first step, we define two classes such as R 1 = Normal subject and R 2 = Patient. The dataset is divided into training data and testing data randomly using 50% data as training data and 50% as testing data.
The training data, with class labels, is used for training of The classification accuracy results using different feature vector size and various classifiersare shown in Fig. 17 and

DISCUSSION
The evaluation of proposed system is carried out using

CONCLUSION
Cerebral malaria is a disease that can be detected and properly treated if an early diagnosis is done. The In summary, we propose a complete computed aided diagnostic and decision support system that is trained on the feature set extracted from the fundus images and very few features provide a robust and accurate assessment of the presence of retinal whitening and the system could be integrated into a clinical diagnosis system to assist ophthalmologists for early diagnosis of cerebral malaria.